<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: @kiwibreaksme</title>
    <description>The latest articles on DEV Community by @kiwibreaksme (@kiwibreaksme).</description>
    <link>https://dev.to/kiwibreaksme</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F865191%2F8d1bfea4-e727-40e2-9d08-a564a58324ce.png</url>
      <title>DEV Community: @kiwibreaksme</title>
      <link>https://dev.to/kiwibreaksme</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/kiwibreaksme"/>
    <language>en</language>
    <item>
      <title>Isometric Seoul: repainting the whole city in 51 minutes, then making it move</title>
      <dc:creator>@kiwibreaksme</dc:creator>
      <pubDate>Tue, 27 Oct 2026 10:14:55 +0000</pubDate>
      <link>https://dev.to/kiwibreaksme/repainting-seoul-in-51-minutes-then-making-it-move-25dn</link>
      <guid>https://dev.to/kiwibreaksme/repainting-seoul-in-51-minutes-then-making-it-move-25dn</guid>
      <description>&lt;p&gt;&lt;em&gt;Originally published at &lt;a href="https://hypeseoul.xyz/making-of/en/" rel="noopener noreferrer"&gt;hypeseoul.xyz/making-of/en&lt;/a&gt;. The map: &lt;a href="https://hypeseoul.xyz" rel="noopener noreferrer"&gt;hypeseoul.xyz&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://hypeseoul.xyz" rel="noopener noreferrer"&gt;HypeSeoul&lt;/a&gt; is an isometric, hand-painted-looking map of Seoul. Under it is a three.js engine that builds the city in 3D from open map data. I fine-tuned an open image model on 33 pairs of "engine render → painting", then used it to repaint all 6,394 tiles of the city (603 km²) on 10 rented A100s in 51 minutes. On top of the painting, a live layer runs cars, boats, people, events and day/night in Seoul time. Some subway lines (1, 2, 3 and Gyeongui–Jungang) are driven by real operating data from the Seoul city government, on a key that allows 1,000 calls a day.&lt;/p&gt;

&lt;p&gt;Things that went wrong are most of this post.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fauhq8s9dmc7urwowb4zn.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fauhq8s9dmc7urwowb4zn.jpg" alt="Seoul, first complete pass" width="800" height="409"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Why
&lt;/h2&gt;

&lt;p&gt;I was browsing and came across &lt;a href="https://isometric.nyc/" rel="noopener noreferrer"&gt;isometric.nyc&lt;/a&gt; by Andy Coenen. It is New York redrawn as an isometric map: real terrain and buildings are rendered, then a fine-tuned image model repaints the render tile by tile. I wanted the same thing for Seoul. It had to look like Seoul from the air, with the Han River and the landmarks readable. And I wanted one more thing. The New York map is beautiful, but nothing moves in it. I wanted cars, trains and people.&lt;/p&gt;

&lt;p&gt;HypeSeoul is an independent project inspired by isometric.nyc. There is no official relationship with Andy Coenen. I followed the method from his public write-up (render the city in 3D, repaint it with a fine-tuned image-edit model, roughly how many training pairs) but reused none of his code: the engine, training and tile tools here are all new. What I wanted to add was:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;a 3D Seoul under the painting, aligned to it pixel for pixel,&lt;/li&gt;
&lt;li&gt;a city that moves on top of it,&lt;/li&gt;
&lt;li&gt;day and night that follow the clock in Seoul.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;It started as the opening screen of a different site. The map kept growing until it became the site.&lt;/p&gt;

&lt;h2&gt;
  
  
  Architecture
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Offline pipeline
&lt;/h3&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt; OSM (buildings, roads, water, parks) ┐
 Overture Maps (missing buildings)    ├─ bake ─&amp;gt; seoul.bin (~585k buildings, terrain, land-use map)
 MOLIT building data (floor counts)   │
 DEM (terrain)                        ┘
            │
            v
 three.js engine ── Playwright captures 1024×1024 tiles (0.3 m/px, 768 px stride, 256 px overlap)
            │
            v
 FLUX.2 klein base 4B + edit LoRA (Modal, 10× A100, 8 steps, guidance 1)
            │
            v
 water_qa → water_fix → forest_grade → harmonize → shade      (automatic fix-up chain)
            │
            ├── landmark / bridge / parking patches (repainted locally, then ECC-aligned)
            v
 build_pyramid: 768 px RGBA WebP, z0–8, alpha = Seoul's boundary
            │                          lights.py: engine buildings → night-light pyramid
            v                                   │
 static object storage &amp;lt;────────────────────────┘
            │
            v
 browser: paint.mjs fetches tiles and composites them over the engine
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  The one thing that holds it together: the same camera
&lt;/h3&gt;

&lt;p&gt;The engine camera is orthographic with a fixed angle (yaw −0.42, pitch 0.56 rad). So a painting only ever translates and scales on screen. Going from world coordinates (x, z, height y) to painting pixels is a single affine map. I measured it in the browser at 0.3 m/px:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="c1"&gt;// engine/preview.mjs — canvas px = anchorPx + AX·dx + AZ·dz + AY·dy (px per metre)&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;AX&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mf"&gt;3.0436&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="mi"&gt;7220&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt; &lt;span class="nx"&gt;AZ&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mf"&gt;1.3592&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mf"&gt;1.6167&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt; &lt;span class="nx"&gt;AY&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="mf"&gt;2.8242&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt; &lt;span class="nx"&gt;Y0&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mf"&gt;18.8516&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;REF&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="mi"&gt;3&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The same constants are copied into the Python tools. Checking painted water against engine water, projecting OSM footprints onto the painting, computing lit windows at night, and putting cars on painted roads all use this one matrix. Almost every fix below works only because the camera never changed.&lt;/p&gt;

&lt;h3&gt;
  
  
  Per-frame passes
&lt;/h3&gt;

&lt;p&gt;&lt;code&gt;render()&lt;/code&gt; in &lt;code&gt;engine/scene.mjs&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="kd"&gt;function&lt;/span&gt; &lt;span class="nf"&gt;render&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="nx"&gt;scene&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;background&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;copy&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;bgDay&lt;/span&gt;&lt;span class="p"&gt;).&lt;/span&gt;&lt;span class="nf"&gt;lerp&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;bgNight&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;U&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;uNight&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;value&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
  &lt;span class="nf"&gt;applyView&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;
  &lt;span class="nf"&gt;renderShadow&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;                                              &lt;span class="c1"&gt;// 1. sun shadow map&lt;/span&gt;
  &lt;span class="nx"&gt;renderer&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;setRenderTarget&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;rt&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt; &lt;span class="nx"&gt;renderer&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;render&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;scene&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;cam&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;   &lt;span class="c1"&gt;// 2. low-res main pass (+depth)&lt;/span&gt;
  &lt;span class="nf"&gt;placePainting&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;                                             &lt;span class="c1"&gt;// 3. painted tiles → 4. moving-things mask&lt;/span&gt;
  &lt;span class="nf"&gt;renderGlow&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;                                                &lt;span class="c1"&gt;// 5. additive light effects&lt;/span&gt;
  &lt;span class="nx"&gt;renderer&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;setRenderTarget&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="kc"&gt;null&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt; &lt;span class="nx"&gt;renderer&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;render&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;postScene&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;postCam&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;   &lt;span class="c1"&gt;// 6. post composite&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt; shadowRT (4096², phones 2048²) ─┐
 rt: low-res colour + depth ─────┼─&amp;gt; POST: depth outlines → AO → night bloom → grading
 paint.rt: tiles (device res) ───┤        → paint composite mix(c, paint, a·(1−mask))
 rtMove: moving-things mask ─────┤        → + glow
 rtGlow: light effects (layer 2) ┘
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Main pass.&lt;/strong&gt; Rendered at screen size divided by &lt;code&gt;pixel&lt;/code&gt; (default 2) and stretched with nearest filtering. That is where the pixel look comes from. Depth goes into a float &lt;code&gt;DepthTexture&lt;/code&gt; that later passes share.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Outlines.&lt;/strong&gt; Second difference of linear depth. It is zero on any plane, whatever the slope, and spikes only at silhouettes. Threshold is &lt;code&gt;max(3, mpp × pixel × 2.4)&lt;/code&gt;, so it scales with zoom.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;AO.&lt;/strong&gt; 8 directions at 3 px and 6 px. Darkens up to 42% where something stands in front.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Deep dive 1: from an image quota to our own model
&lt;/h2&gt;

&lt;h3&gt;
  
  
  The quota wall
&lt;/h3&gt;

&lt;p&gt;The first version repainted engine renders with GPT image generation, tile by tile (1024 px, 0.3 m/px, 768 px stride, with a strip of the already-painted neighbour attached so edges would match). Yeouido alone looked great. Two problems. GPT moved the Han River, added ponds that do not exist, and put towers in parks. And I hit the image limit on my plan at about 348 images in a day. Seoul is about 6,000 tiles, plus night.&lt;/p&gt;

&lt;p&gt;Lesson one, which cost me two days: &lt;strong&gt;before any batch generation, multiply tiles × cost per image × quota × days.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;I also tried laying painted fragments as shader patterns per zoning type. As long as the 3D boxes showed underneath, it never read as a painting. What I actually needed was a model that had learned the style.&lt;/p&gt;

&lt;h3&gt;
  
  
  Training
&lt;/h3&gt;

&lt;p&gt;I trained an edit LoRA on FLUX.2 klein 4B (Apache 2.0). Input is the engine render, target is a painting of the same tile.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fselk1jmfmyaova0nmsa9.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fselk1jmfmyaova0nmsa9.jpg" alt="Satellite / engine render / base model before training" width="800" height="584"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Run&lt;/th&gt;
&lt;th&gt;Hardware, tool&lt;/th&gt;
&lt;th&gt;Pairs&lt;/th&gt;
&lt;th&gt;Settings&lt;/th&gt;
&lt;th&gt;Speed&lt;/th&gt;
&lt;th&gt;Outcome&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;r1&lt;/td&gt;
&lt;td&gt;MacBook M3 Pro, mflux&lt;/td&gt;
&lt;td&gt;92&lt;/td&gt;
&lt;td&gt;768 px, rank 16, 8-bit, 1,104 steps&lt;/td&gt;
&lt;td&gt;77 s/step&lt;/td&gt;
&lt;td&gt;~4 nights projected. Stopped after finding input/target mismatch&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;r2&lt;/td&gt;
&lt;td&gt;MacBook&lt;/td&gt;
&lt;td&gt;33&lt;/td&gt;
&lt;td&gt;512 px, 396 steps&lt;/td&gt;
&lt;td&gt;57 s/step&lt;/td&gt;
&lt;td&gt;Invented water on the distilled model. Undertrained&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;r3&lt;/td&gt;
&lt;td&gt;Kaggle 2× T4 (free), ai-toolkit&lt;/td&gt;
&lt;td&gt;33&lt;/td&gt;
&lt;td&gt;512 px reference, 1,000 steps, rank 32/64&lt;/td&gt;
&lt;td&gt;31–35 s/step&lt;/td&gt;
&lt;td&gt;Coarse detail; canals and forest islands remained&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;r4&lt;/td&gt;
&lt;td&gt;Modal A100-80GB, ai-toolkit&lt;/td&gt;
&lt;td&gt;33&lt;/td&gt;
&lt;td&gt;1024 px ref and target, no quantization, rank 64, 1,500 steps&lt;/td&gt;
&lt;td&gt;2.41 s/step&lt;/td&gt;
&lt;td&gt;Shipped&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Where it got stuck:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The training data was wrong.&lt;/strong&gt; I warped satellite imagery (used for review only, never as model input) into the camera angle and compared. Where OSM was missing buildings, the target paintings had filled the empty land with forest. The model was learning "empty lot = forest". Worse, between captures the engine had started filling empty blocks with placeholder buildings, so inputs and targets no longer matched. &lt;code&gt;pairscore.py&lt;/code&gt; kept only pairs where fewer than 6% of real buildings were missing from the input and less than 20% of the target was invented. That left 33 pairs.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The model kept inventing water.&lt;/strong&gt; A LoRA trained on the base model, when plugged into the fast distilled model, turned wide roads into canals. Same result on all four random seeds, so it was learned behaviour, not noise. Run on the base model, the canals went away. Final inference uses the same &lt;code&gt;FLUX.2-klein-base-4B&lt;/code&gt; it was trained on.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;OOM on a T4, always at step 3.&lt;/strong&gt; fp16 went NaN (a known FLUX-family issue), fp32 was as slow as the laptop, and bf16 died at step 3 every time, through defragmentation, rank 16 and CPU offload. Reading the trainer's code, I found it fed the edit reference at its original size: a 1024 px reference against a 512 px target, 4× the tokens. A 512 px reference finished at 27 s/step. &lt;strong&gt;If it dies at the same step every time, look at input size before settings.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Rank and steps.&lt;/strong&gt; r4 trained rank 32 and rank 64 in parallel for 1,500 steps (about 65 minutes). Rank 32 painted canals and ponds at 500, 1,000 and 1,500 steps. Rank 64 still had them at 1,000 and settled at 1,500. On 16 held-out tiles: zero invented water. On 3 test tiles × 2 seeds: 0/6.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight yaml"&gt;&lt;code&gt;&lt;span class="c1"&gt;# ai-toolkit config generated by train_modal.py (excerpt)&lt;/span&gt;
&lt;span class="na"&gt;network&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="pi"&gt;{&lt;/span&gt;&lt;span class="nv"&gt;type&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="nv"&gt;lora&lt;/span&gt;&lt;span class="pi"&gt;,&lt;/span&gt; &lt;span class="nv"&gt;linear&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="nv"&gt;64&lt;/span&gt;&lt;span class="pi"&gt;,&lt;/span&gt; &lt;span class="nv"&gt;linear_alpha&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="nv"&gt;32&lt;/span&gt;&lt;span class="pi"&gt;}&lt;/span&gt;
&lt;span class="na"&gt;train&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="pi"&gt;{&lt;/span&gt;&lt;span class="nv"&gt;batch_size&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="nv"&gt;1&lt;/span&gt;&lt;span class="pi"&gt;,&lt;/span&gt; &lt;span class="nv"&gt;steps&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="nv"&gt;1500&lt;/span&gt;&lt;span class="pi"&gt;,&lt;/span&gt; &lt;span class="nv"&gt;lr&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="nv"&gt;0.0001&lt;/span&gt;&lt;span class="pi"&gt;,&lt;/span&gt; &lt;span class="nv"&gt;optimizer&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="nv"&gt;adamw8bit&lt;/span&gt;&lt;span class="pi"&gt;,&lt;/span&gt;
        &lt;span class="nv"&gt;gradient_checkpointing&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="nv"&gt;true&lt;/span&gt;&lt;span class="pi"&gt;,&lt;/span&gt; &lt;span class="nv"&gt;dtype&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="nv"&gt;bf16&lt;/span&gt;&lt;span class="pi"&gt;,&lt;/span&gt; &lt;span class="nv"&gt;noise_scheduler&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="nv"&gt;flowmatch&lt;/span&gt;&lt;span class="pi"&gt;}&lt;/span&gt;
&lt;span class="na"&gt;model&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="pi"&gt;{&lt;/span&gt;&lt;span class="nv"&gt;arch&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="nv"&gt;flux2_klein_4b&lt;/span&gt;&lt;span class="pi"&gt;,&lt;/span&gt; &lt;span class="nv"&gt;name_or_path&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="nv"&gt;black-forest-labs/FLUX.2-klein-base-4B&lt;/span&gt;&lt;span class="pi"&gt;,&lt;/span&gt;
        &lt;span class="nv"&gt;quantize&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="nv"&gt;false&lt;/span&gt;&lt;span class="pi"&gt;,&lt;/span&gt; &lt;span class="nv"&gt;quantize_te&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="nv"&gt;true&lt;/span&gt;&lt;span class="pi"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Operational mistakes.&lt;/strong&gt; A resume script that found no checkpoint returned the directory listing instead, and the first overnight run died after one second; I found out three hours later. Later, loading the finished LoRA just to sample, ai-toolkit ignored the saved step count and retrained from step 0, and killing the local process did not stop the remote app. Since then every sampling run, including the full city, is a "training job that is already finished":&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="c1"&gt;# convert: sample only, no training (start_step = steps) — simplified
&lt;/span&gt;&lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nf"&gt;run&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;name&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;steps&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;rank&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;1024&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;100000&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;W&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;/vol/conv&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;lora&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sa"&gt;f&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;, pretrained_lora_path: &lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;lora_path&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
           &lt;span class="n"&gt;items&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;items&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;extra_train&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sa"&gt;f&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="se"&gt;\n&lt;/span&gt;&lt;span class="s"&gt;        skip_first_sample: true&lt;/span&gt;&lt;span class="se"&gt;\n&lt;/span&gt;&lt;span class="s"&gt;        start_step: &lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;steps&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpduxsic1s4l7ee4f4pmv.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpduxsic1s4l7ee4f4pmv.jpg" alt="r4 final comparison sheet" width="800" height="275"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  The night run: 51 minutes
&lt;/h3&gt;

&lt;p&gt;I timed 6 tiles × 4 settings on an A100:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Setting&lt;/th&gt;
&lt;th&gt;Per tile&lt;/th&gt;
&lt;th&gt;Notes&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;8 steps, guidance 3&lt;/td&gt;
&lt;td&gt;8.6 s&lt;/td&gt;
&lt;td&gt;baseline&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;8 steps, guidance 1&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;4.1 s&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;ground and ripples slightly flatter; structure, buildings, colour nearly identical&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;6 steps, guidance 3&lt;/td&gt;
&lt;td&gt;5.8 s&lt;/td&gt;
&lt;td&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;6 steps, guidance 1&lt;/td&gt;
&lt;td&gt;3.2 s&lt;/td&gt;
&lt;td&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;4 steps, guidance 1&lt;/td&gt;
&lt;td&gt;2.3 s&lt;/td&gt;
&lt;td&gt;ground goes soft&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Guidance 1 turns classifier-free guidance off. At the same 8 steps it more than halved the time per tile. None of the settings invented water. At 8 steps and guidance 1, 6,000 tiles came to about 6.8 hours of GPU, roughly $17, inside my remaining free credit.&lt;/p&gt;

&lt;p&gt;The run: 01:01, engine capture starts (headless Chromium, 1024×1024 viewport, UI, trees and the live layer hidden, about 0.7 s per tile). 02:37, 10 A100 containers start converting, 628–629 tiles each, 4.9 s per tile. 51 minutes later all 6,394 tiles were done. Zero errors, zero OOM, zero NaN, zero missing tiles.&lt;/p&gt;

&lt;p&gt;Total Modal usage up to that point was $35.88, including $30 of free credit. It was above my $31 estimate because each container loads the model before it starts.&lt;/p&gt;

&lt;h3&gt;
  
  
  Making 6,394 paintings into one
&lt;/h3&gt;

&lt;p&gt;Each tile is painted independently. The fix-up chain makes them one image.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Water QA.&lt;/strong&gt; Because engine and painting share pixels, I can ask: where the engine has water, is the painting blue? A tile fails if fewer than 85% of engine-water pixels satisfy &lt;code&gt;B &amp;gt; R + 30 &amp;amp;&amp;amp; B &amp;gt; G&lt;/code&gt;. Of 949 tiles with water, 194 were at least 80% water, and 176 of those failed. Tiles in the middle of the Han, with no bank in view, were painted as autumn forest. The training set had no water-only tiles.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fg0s2jkfe9nj4zptfcjid.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fg0s2jkfe9nj4zptfcjid.jpg" alt="Water-only tiles painted as forest" width="800" height="267"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The fix rebuilds water in three layers: one city-wide blue, a slow brightness field in canvas coordinates (so it continues across tiles; the first version used per-tile references and the river showed squares), and ripples, taken from the tile itself if it kept its water, procedural otherwise.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="c1"&gt;# water_fix.py (excerpt)
&lt;/span&gt;&lt;span class="n"&gt;det&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;OWN&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;o&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="nf"&gt;lp&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;o&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;6&lt;/span&gt;&lt;span class="p"&gt;))&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;blue&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;=&lt;/span&gt; &lt;span class="mf"&gt;0.85&lt;/span&gt; &lt;span class="k"&gt;else&lt;/span&gt; &lt;span class="nf"&gt;ripples&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;t&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;new&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;np&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;clip&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;PM&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="nf"&gt;field&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;t&lt;/span&gt;&lt;span class="p"&gt;)[...,&lt;/span&gt; &lt;span class="bp"&gt;None&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="n"&gt;det&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;255&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt; &lt;span class="n"&gt;out&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;o&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;a&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="n"&gt;new&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="n"&gt;a&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Colour seams (harmonize).&lt;/strong&gt; Each tile's colour drifts a little, so the grid showed. Unknowns: log gains at 3×3 control points per tile and channel. Constraints: neighbouring overlap strips should match, plus terms pulling gains to zero and keeping them smooth. One sparse least-squares solve (&lt;code&gt;lsqr&lt;/code&gt;) at 1/8 scale, gains upsampled bilinearly. Overlap differences dropped to about a quarter.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Ghosting (build_pyramid).&lt;/strong&gt; At first I faded linearly across the full 256 px overlap. Where two tiles drew the same building slightly differently, you got a translucent double. Now each tile owns the central 768 px cell and blending happens only in 96 px around each cell border:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;ramp&lt;/span&gt;&lt;span class="p"&gt;():&lt;/span&gt;  &lt;span class="c1"&gt;# weight of this tile in cell-local coords -R/2..C+R/2 (R=96, C=768)
&lt;/span&gt;    &lt;span class="n"&gt;u&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;np&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;arange&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="n"&gt;R&lt;/span&gt; &lt;span class="o"&gt;//&lt;/span&gt; &lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;C&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="n"&gt;R&lt;/span&gt; &lt;span class="o"&gt;//&lt;/span&gt; &lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;dtype&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;np&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;float32&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="n"&gt;np&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;clip&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;np&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;minimum&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;u&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="n"&gt;R&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;C&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="n"&gt;R&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mi"&gt;2&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;u&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="n"&gt;R&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Alpha comes from Seoul's real administrative boundary, projected with the same camera and terrain, feathered over 160 px. The map ends in the shape of Seoul, not a staircase of tiles.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Long things break across tiles.&lt;/strong&gt; The same Han bridge sat up to 40+ px apart in neighbouring tiles, with different widths and lane designs. Shifting tiles failed. Normalising deck width and sharpening the blend near bridges (&lt;code&gt;w ** (1 + 3 * bridge_mask)&lt;/code&gt;) removed ghosts but not the design steps. What worked: repaint each bridge bank to bank in one image. 24 crossings; the 15 long ones split into two overlapping images, the second continuing from the first. 37 images.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Favoy8k6292dzuuhbitwr.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Favoy8k6292dzuuhbitwr.jpg" alt="Repainted Han River bridges" width="800" height="519"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Landmarks, which the model painted as generic buildings, got the same treatment. &lt;code&gt;lm_prep.py&lt;/code&gt; projects the OSM footprint and height onto the painting in red (minus anything occluded by buildings in front, or the repaint erases them), and only the marked area is repainted. &lt;code&gt;lm_apply.py&lt;/code&gt; aligns the result with ECC affine registration on pixels well outside the mark and saves a feathered patch in world coordinates. Alignment error stayed under about 0.6 px.&lt;/p&gt;

&lt;p&gt;Yeouido taught the general rule. My carefully painted per-tile GPT version stepped at every tile border. After days of fixes I switched it off and used the model's seamless painting, repainting only the 63 Building, IFC, Parc1 and Yeouido Park on top. A continuous whole beats pretty fragments.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fj6lr0ngsxoel084gpw0v.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fj6lr0ngsxoel084gpw0v.jpg" alt="Yeouido: per-tile GPT mosaic vs the fine-tuned model" width="800" height="768"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  A lot of model mistakes were data mistakes
&lt;/h3&gt;

&lt;p&gt;Low buildings kept getting painted as apartment towers. Only 92k of ~585k buildings (16%) had a real height; the estimates pushed wide, low markets and schools into tower-sized blocks, and the model painted them, faithfully, as apartments. I joined the national GIS building dataset from Korea's Ministry of Land, Infrastructure and Transport (MOLIT), which carries attributes from the building register (건축물대장, the official record of every building). Matching "30% of our footprint" turned a 15 m shop into 250 m; requiring 50% overlap &lt;em&gt;in both directions&lt;/em&gt;, and keeping real OSM heights, mostly worked (one more rule caught a podium that had inherited a whole tower site's height). 204k buildings got official heights (the 63 Building went from 37 m to 250 m). I repainted the 1,747 tiles whose wall area changed by more than 4%, and southwest Seoul went back from 15–25-storey blocks to 3–5-storey alleys.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F3dib9m4r0zvpcajlitix.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F3dib9m4r0zvpcajlitix.jpg" alt="Before and after the official floor counts: tower blocks back to low-rise villas" width="760" height="1034"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;A pre-launch catch: the 63 Building.&lt;/strong&gt; A last check found a giant gold podium painted over the mid-rise blocks in front of the 63 Building. The cause was data. When I filled missing heights from the building register, the whole complex outline (115×133 m) inherited the tower's 249 m, and I had marked that outline in red. GPT did as told and filled it. The fix was a two-colour mark: red for what is real (the tower at 250 m and a 14 m podium), blue for things the model had drawn from bad data, with the instruction "no building here, continue the riverside park behind". GPT ignored the blue the first time and kept the grey slab. So I pre-filled the blue area of the reference image with OpenCV inpainting, which left nothing to copy. Buildings standing in front are cut out of the mark using the engine's building heights. IFC had the opposite problem: the old painted tower was wider than the mark, so its flat roof ghosted beside the new faceted spire. Pre-filling the whole mark in the reference (&lt;code&gt;prefill=True&lt;/code&gt;) removed anything to copy. Lesson: the mark is exactly what gets overwritten, so look at the surroundings first.&lt;/p&gt;

&lt;h2&gt;
  
  
  Deep dive 2: painted tiles under a live 3D city
&lt;/h2&gt;

&lt;p&gt;The painting covers the engine's static city. But cars, people and events have to stay visible on top of it and still disappear behind buildings.&lt;/p&gt;

&lt;p&gt;The moving things are drawn a second time into a mask, in white. That pass shares the main pass's depth texture, with depth test on and depth write off. A car behind a building is hidden in the mask too, so the painting stays there.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="c1"&gt;// mask material: same transform as the engine shader, nudged toward the camera (simplified)&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;moveMat&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nx"&gt;THREE&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;ShaderMaterial&lt;/span&gt;&lt;span class="p"&gt;({&lt;/span&gt; &lt;span class="na"&gt;side&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nx"&gt;THREE&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;DoubleSide&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="na"&gt;depthWrite&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="na"&gt;depthFunc&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nx"&gt;THREE&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;LessEqualDepth&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="na"&gt;vertexShader&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;`... gl_Position.z -= .0004 * gl_Position.w; }`&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="na"&gt;fragmentShader&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;void main() { gl_FragColor = vec4(1.0); }&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt; &lt;span class="p"&gt;});&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight glsl"&gt;&lt;code&gt;&lt;span class="c1"&gt;// POST_FS: the painting covers only where the mask is empty&lt;/span&gt;
&lt;span class="kt"&gt;vec4&lt;/span&gt; &lt;span class="n"&gt;p&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;texture2D&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;tPaint&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;vUv&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
&lt;span class="kt"&gt;float&lt;/span&gt; &lt;span class="n"&gt;mv&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;texture2D&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;tMove&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;vUv&lt;/span&gt;&lt;span class="p"&gt;).&lt;/span&gt;&lt;span class="n"&gt;r&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="n"&gt;c&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;mix&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;c&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;p&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;rgb&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="n"&gt;max&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;p&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;a&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mf"&gt;1e-3&lt;/span&gt;&lt;span class="p"&gt;),&lt;/span&gt; &lt;span class="n"&gt;p&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;a&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;mv&lt;/span&gt;&lt;span class="p"&gt;));&lt;/span&gt;
&lt;span class="n"&gt;c&lt;/span&gt; &lt;span class="o"&gt;+=&lt;/span&gt; &lt;span class="n"&gt;texture2D&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;tGlow&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;vUv&lt;/span&gt;&lt;span class="p"&gt;).&lt;/span&gt;&lt;span class="n"&gt;rgb&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The moving things themselves come from the low-res main pass, so they get the same pixel size, the same depth outlines, the same AO and the same sun and night colour as the buildings. That is why a car does not look pasted on. Hold on to that; it matters in deep dive 4.&lt;/p&gt;

&lt;p&gt;Two problems on the way:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Additive light effects&lt;/strong&gt; (searchlights, fireworks) were caught by the mask. Wherever a beam passed, the painting was cut away and the bare engine showed through. I moved light effects to camera layer 2, rendered them into their own &lt;code&gt;rtGlow&lt;/code&gt;, and added them last. Now a beam brightens the painting instead of punching a hole in it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Trains on curves&lt;/strong&gt; looked jagged when each car was rotated by the track angle at one point. Each car now stands on two bogie points:
&lt;/li&gt;
&lt;/ul&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="c1"&gt;// each car stands on front/back bogies (±8 m); cars 20 m apart&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;dc&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nx"&gt;t&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;d&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="nx"&gt;t&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;dir&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="nx"&gt;c&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;20&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="nx"&gt;run&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;at&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;dc&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="mi"&gt;8&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;f&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt; &lt;span class="nx"&gt;run&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;at&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;dc&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="mi"&gt;8&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;b&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;x&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;f&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;x&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="nx"&gt;b&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;x&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;z&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;f&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;z&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="nx"&gt;b&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;z&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;a&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nb"&gt;Math&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;atan2&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;f&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;x&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="nx"&gt;b&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;x&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;f&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;z&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="nx"&gt;b&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;z&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Night is data, not another painting.&lt;/strong&gt; Repainting night tiles would double the cost. Instead the browser tints the painting dark and &lt;code&gt;lights.py&lt;/code&gt; adds windows computed from engine buildings with the same camera, with whole floors switched off by a deterministic hash. 9,704 tiles, 6 minutes locally, $0. One trap: after I repainted Namsan Tower, the night layer kept lighting windows in the office box the model had first drawn under it. Fix the painting and you must fix what was derived from it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Performance.&lt;/strong&gt; Loading all of Seoul at once took 10–15 s to the first frame, with a median frame of 18 ms and spikes of 70–100 ms. The GPU was drawing the full 3D city every frame under the painting: 26.9M triangles and 1.7M instances (450k trees alone were 9.1M triangles). three.js draws every instance of an &lt;code&gt;InstancedMesh&lt;/code&gt;. &lt;code&gt;engine/cull.mjs&lt;/code&gt; sorts instances into 400 m ground cells ahead of time and, whenever the visible cell range changes, copies only visible cells to the front of the buffer:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="k"&gt;for &lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="kd"&gt;let&lt;/span&gt; &lt;span class="nx"&gt;cx&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nx"&gt;cx0&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="nx"&gt;cx&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;=&lt;/span&gt; &lt;span class="nx"&gt;cx1&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="nx"&gt;cx&lt;/span&gt;&lt;span class="o"&gt;++&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;for &lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="kd"&gt;let&lt;/span&gt; &lt;span class="nx"&gt;cz&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nx"&gt;cz0&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="nx"&gt;cz&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;=&lt;/span&gt; &lt;span class="nx"&gt;cz1&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="nx"&gt;cz&lt;/span&gt;&lt;span class="o"&gt;++&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;r&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nx"&gt;range&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;get&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;`&lt;/span&gt;&lt;span class="p"&gt;${&lt;/span&gt;&lt;span class="nx"&gt;cx&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="s2"&gt;,&lt;/span&gt;&lt;span class="p"&gt;${&lt;/span&gt;&lt;span class="nx"&gt;cz&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="s2"&gt;`&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt; &lt;span class="k"&gt;if &lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;!&lt;/span&gt;&lt;span class="nx"&gt;r&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;continue&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="nx"&gt;dm&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;set&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;SM&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;subarray&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;r&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;16&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;r&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;16&lt;/span&gt;&lt;span class="p"&gt;),&lt;/span&gt; &lt;span class="nx"&gt;k&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;16&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;   &lt;span class="c1"&gt;// one cell = one contiguous range&lt;/span&gt;
  &lt;span class="nx"&gt;k&lt;/span&gt; &lt;span class="o"&gt;+=&lt;/span&gt; &lt;span class="nx"&gt;r&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="nx"&gt;r&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;];&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="nx"&gt;mesh&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;count&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nx"&gt;k&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;   &lt;span class="c1"&gt;// + addUpdateRange(0, k * 16) uploads only the front&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Same Yeouido view: 1.7M instances to about 5k, JS memory 2.2 GB to 0.75 GB, 46 to 73 fps. Skipping alley road surfaces under the painting took triangles to 4.2M, and pre-baking parked-car, pedestrian and street-tree positions in Node with the same worker code cut city build time from about 7 s to 1.25–1.5 s.&lt;/p&gt;

&lt;h2&gt;
  
  
  Deep dive 3: real subway positions on 1,000 calls a day
&lt;/h2&gt;

&lt;p&gt;The original subway was fake. Above-ground track segments joined together, one train every 2.6 km, random speed between 16 and 22 m/s, hidden in tunnels. Right before launch I replaced it with real operating data.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The data.&lt;/strong&gt; The Seoul Open Data Plaza has a real-time train position API. It does not give coordinates. It says which train entered, arrived at, or departed which station: train number, station name, direction (up/down; inner/outer loop on Line 2), status, and timestamp. One key allows &lt;strong&gt;1,000 calls per day&lt;/strong&gt;, and the terms forbid sharing or exposing the key.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Which lines.&lt;/strong&gt; Only above-ground track is visible on the map, so I measured it per line (Line 1 family 195 km, Gyeongui–Jungang 99 km, Line 2 31 km, down to 9 km). Lines 1, 2, 3 and Gyeongui–Jungang get real data; the rest stay simulated.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Visitors never call the city.&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;visitor browser ──(every 5 min)──▶ /api/subway  (Vercel function + CDN cache, 10 min)
                                       │  once per 10 min
                                       ▼
                           Seoul real-time train position API (4 lines = 4 calls)
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The key lives only in a server environment variable. The function sets &lt;code&gt;s-maxage=600&lt;/code&gt;, so the CDN serves the same copy for 10 minutes. However many visitors there are, the city API is called 4 times per 10 minutes per region: 576 calls a day.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="c1"&gt;// api/subway.js (simplified)&lt;/span&gt;
&lt;span class="nx"&gt;res&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;setHeader&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;cache-control&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;public, max-age=60, s-maxage=600, stale-while-revalidate=300&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If the quota runs out or the city returns an empty response, the function keeps serving the last good data. If there is none, the client falls back to simulated trains for that line only. Trains never vanish. One catch: CDN copies are per region, so overseas traffic multiplies calls by the number of regions. The plan is to register the project as a use case with the open data portal to raise the limit.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Station names to map points.&lt;/strong&gt; The API gives names only. I tried the Overpass API for station coordinates; it timed out twice. So I pulled 524 station nodes directly from the OSM extract (&lt;code&gt;.pbf&lt;/code&gt;) the map was built from, projected with the map's own formula:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;proj&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;lat&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;lon&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="p"&gt;[(&lt;/span&gt;&lt;span class="nx"&gt;lon&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="nx"&gt;LON0&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="nx"&gt;KX&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;lat&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="nx"&gt;LAT0&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="nx"&gt;KZ&lt;/span&gt;&lt;span class="p"&gt;];&lt;/span&gt;   &lt;span class="c1"&gt;// same origin and scale as the map&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;When a name matches several stations, I take the candidate closest to that line's above-ground track. A station more than 220 m from the track is treated as underground, and trains there are not drawn. They are in the tunnel.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Which way is "up"?&lt;/strong&gt; Nothing says which screen direction along the track the API's up/down means. I took two snapshots two minutes apart (8 calls) and collected the 97 trains that moved from station A to station B. That gives samples of "this line, this direction, near here, moves this way". At runtime, for a train at a station, I find the nearest learnt sample for the same line and direction, and take the sign of its dot product with the track tangent.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="c1"&gt;// engine/realtrains.mjs (simplified)&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;dir&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nb"&gt;Math&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;sign&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nf"&gt;tangentAt&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;station&lt;/span&gt;&lt;span class="p"&gt;).&lt;/span&gt;&lt;span class="nf"&gt;dot&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nf"&gt;nearestLearntMove&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;line&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;updn&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;station&lt;/span&gt;&lt;span class="p"&gt;)));&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;off&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="mi"&gt;250&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;30&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;3&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="mi"&gt;800&lt;/span&gt; &lt;span class="p"&gt;}[&lt;/span&gt;&lt;span class="nx"&gt;status&lt;/span&gt;&lt;span class="p"&gt;];&lt;/span&gt;   &lt;span class="c1"&gt;// entering, arrived, departed, left previous station&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;d&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nx"&gt;station&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;d&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="nx"&gt;dir&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;off&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="nb"&gt;Math&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;min&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;7000&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;secondsSinceReport&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;11&lt;/span&gt;&lt;span class="p"&gt;));&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Filling the 10 minutes.&lt;/strong&gt; Data can be up to 10 minutes old. From the report time, each train keeps moving at 11 m/s, the average subway speed including dwell time. When new data arrives, trains with the same number glide to their new position over 6 seconds instead of teleporting. Real trains take over the line's instanced-mesh slots; leftover slots are parked off-screen at y = −999.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Deployment.&lt;/strong&gt; With a static setup (&lt;code&gt;outputDirectory: "."&lt;/code&gt;, empty build command), files in &lt;code&gt;api/&lt;/code&gt; did not become functions and returned 404. Declaring &lt;code&gt;functions&lt;/code&gt; in &lt;code&gt;vercel.json&lt;/code&gt; fixed it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The honest result.&lt;/strong&gt; Right after deploy, 155 trains were running and 45 of them were above ground and drawn. The fake version had about 140 trains spread evenly. Real data made the map emptier, because most trains are underground at any moment. I have not decided yet between showing it honestly and making it look busy. Two bugs showed up immediately: a 3 km cap on dead-reckoning made trains stop mid-track between updates (raised to 7 km), and trains that ran past the end of the track piled up there (now hidden).&lt;/p&gt;

&lt;p&gt;It is not real-time in the strict sense. Data is up to 10 minutes old and positions in between are predicted. So the site says "driven by real operating data", not "real-time subway".&lt;/p&gt;

&lt;h2&gt;
  
  
  Deep dive 4: mobile GPU memory and the flicker
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Resolution.&lt;/strong&gt; The renderer's &lt;code&gt;pixelRatio&lt;/code&gt; was 1, so Retina screens and phones got a 1× image stretched up. The final output and the paint layer now render at device resolution (up to 2×); the engine main pass stays low-res for the pixel look. On a 6K monitor, 2× means 24 Mpx per frame, and the night view once came out black, so output is capped at about 8.3 Mpx:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="nx"&gt;DPR&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nb"&gt;Math&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;max&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nb"&gt;Math&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;min&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;devicePixelRatio&lt;/span&gt; &lt;span class="o"&gt;||&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nb"&gt;Math&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;sqrt&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mf"&gt;8.3e6&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;w&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="nx"&gt;h&lt;/span&gt;&lt;span class="p"&gt;))));&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Flicker.&lt;/strong&gt; Stepping through an iPhone screen recording frame by frame: at night, painted tiles flickered square by square, then the whole painting vanished and only the engine was left. I read that as GPU memory exhaustion. Three things stacked up. At 2× the loader picked finer levels, so the cache filled faster. Night-light textures were not counted in the budget at all. And mipmaps added about 33% to every texture.&lt;/p&gt;

&lt;p&gt;A 768² RGBA8 tile is about 2.4 MB, about 3.1 MB with mipmaps (back-of-envelope, not measured). The old phone budget of 72 tiles was about 227 MB for paintings alone, with lights potentially doubling that at night. The new budget is 56 tiles including lights, no mipmaps, about 132 MB.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;COARSE&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;matchMedia&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;(pointer: coarse)&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;).&lt;/span&gt;&lt;span class="nx"&gt;matches&lt;/span&gt; &lt;span class="o"&gt;||&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nc"&gt;URLSearchParams&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;location&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;search&lt;/span&gt;&lt;span class="p"&gt;).&lt;/span&gt;&lt;span class="nf"&gt;has&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;coarse&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;MAX_PYR&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nx"&gt;COARSE&lt;/span&gt; &lt;span class="p"&gt;?&lt;/span&gt; &lt;span class="mi"&gt;56&lt;/span&gt; &lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;140&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="c1"&gt;// upload: no mipmaps on phones (levels are chosen per device pixel, so little minification)&lt;/span&gt;
&lt;span class="nx"&gt;tx&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;minFilter&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nx"&gt;COARSE&lt;/span&gt; &lt;span class="p"&gt;?&lt;/span&gt; &lt;span class="nx"&gt;THREE&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;LinearFilter&lt;/span&gt; &lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nx"&gt;THREE&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;LinearMipmapLinearFilter&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="nx"&gt;tx&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;generateMipmaps&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="o"&gt;!&lt;/span&gt;&lt;span class="nx"&gt;COARSE&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;I checked it on desktop with &lt;code&gt;?coarse&lt;/code&gt; to emulate the phone budget, panning the night view in 40 steps and logging how much of the screen the painting covered: 0.8–1 while panning, 1 at rest. I have not re-checked on a real iPhone yet.&lt;/p&gt;

&lt;p&gt;The loader around it matters as much. Missing tiles are covered by the nearest level already shown, requests are queued nearest-to-centre 8 at a time and aborted if stale, decoding runs through &lt;code&gt;createImageBitmap&lt;/code&gt;, and at most 3 textures go to the GPU per frame.&lt;/p&gt;

&lt;h2&gt;
  
  
  Numbers
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Item&lt;/th&gt;
&lt;th&gt;Value&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Tiles&lt;/td&gt;
&lt;td&gt;6,394 (603 km²) in the first pass; 6,470 after extent and terrain fixes&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Tile spec&lt;/td&gt;
&lt;td&gt;1024 px, 0.3 m/px, 768 px stride (256 px overlap), 96 px border blend&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Training data&lt;/td&gt;
&lt;td&gt;33 filtered pairs (engine render ↔ painting)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Model&lt;/td&gt;
&lt;td&gt;FLUX.2 klein base 4B edit LoRA, rank 64 (alpha 32), 1,500 steps, lr 1e-4, bf16&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Training speed&lt;/td&gt;
&lt;td&gt;MacBook 57–77 s/step → Kaggle T4 27 s → A100 2.41 s&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Inference&lt;/td&gt;
&lt;td&gt;8 steps, guidance 1; A100 4.1 s/tile (test), 4.9 s/tile (full run)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Full city conversion&lt;/td&gt;
&lt;td&gt;10 A100s in parallel, 51 minutes, 0 errors&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GPU spend&lt;/td&gt;
&lt;td&gt;Modal ~$36 to first complete map (incl. $30 free credit); $44.98 cumulative as of Oct 23&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Buildings&lt;/td&gt;
&lt;td&gt;~585k, 204k with official heights&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Han River bridges&lt;/td&gt;
&lt;td&gt;24 crossings, 37 repainted images&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Night lights&lt;/td&gt;
&lt;td&gt;9,704 tiles, 6 min locally, $0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Tiles served&lt;/td&gt;
&lt;td&gt;768 px WebP z0–8, ~2.0 GB + 439 MB lights, 22,186 files&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Render&lt;/td&gt;
&lt;td&gt;instances 1.7M → ~5k, triangles 26.9M → 4.2M, 46 → 73 fps, JS memory 2.2 GB → 0.75 GB&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Loading&lt;/td&gt;
&lt;td&gt;worker ~7 s → 1.25–1.5 s; first painted frame under 2 s&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GPU texture budget&lt;/td&gt;
&lt;td&gt;desktop 140; phone 72 → 56 (lights included, no mipmaps)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Subway&lt;/td&gt;
&lt;td&gt;4 real lines, 576 API calls/day, 155 trains running / 45 drawn at deploy&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Characters&lt;/td&gt;
&lt;td&gt;36 from Korean folklore&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h2&gt;
  
  
  What I'd do differently
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Multiply before generating.&lt;/strong&gt; Tiles × cost × quota × days. Not doing it cost me two days.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Suspect the input first.&lt;/strong&gt; Forest on empty lots and villas turned into towers were data problems, not model problems.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Same failure at the same step means input size.&lt;/strong&gt; Reading the trainer's source was faster than any tuning.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Don't trust your training tool's state.&lt;/strong&gt; It ignored a loaded LoRA's step count and retrained; the remote app kept running after the local process died.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A continuous whole beats pretty fragments.&lt;/strong&gt; Anything long (riverbanks, bridges) gets repainted in one image.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Keep the camera fixed and many things come for free:&lt;/strong&gt; automatic QA, alignment, night lights, occlusion, click targets.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Put every texture in the GPU budget.&lt;/strong&gt; One missing texture type and the hidden 33% of mipmaps took the phone down.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Decide how honest live data will look before wiring it up.&lt;/strong&gt; Real trains made the map emptier.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Credits and data sources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Inspired by &lt;a href="https://isometric.nyc/" rel="noopener noreferrer"&gt;isometric.nyc&lt;/a&gt; by Andy Coenen. HypeSeoul is an independent project with no official relationship.&lt;/li&gt;
&lt;li&gt;© OpenStreetMap contributors, available under the Open Database License (ODbL)&lt;/li&gt;
&lt;li&gt;Overture Maps Foundation buildings theme (ODbL), which includes OpenStreetMap, Microsoft Global ML Building Footprints (ODbL), and Google Open Buildings, Esri Community Maps and Shi et al. East Asian buildings (Zenodo 8174931) under CC BY 4.0&lt;/li&gt;
&lt;li&gt;Korea Ministry of Land, Infrastructure and Transport (MOLIT), GIS building dataset (building register / 건축물대장 attributes), CC BY&lt;/li&gt;
&lt;li&gt;Terrain: AWS Open Data Terrain Tiles (Terrarium; Mapzen Joerd, built from SRTM and other public elevation data)&lt;/li&gt;
&lt;li&gt;Seoul Metropolitan Government public data (Seoul Open Data Plaza, real-time subway positions)&lt;/li&gt;
&lt;li&gt;Satellite imagery was used only for visual review, never as model input (Esri World Imagery)&lt;/li&gt;
&lt;li&gt;Model: FLUX.2 klein 4B (Apache 2.0); training with ai-toolkit (MIT, Ostris); rendering with three.js&lt;/li&gt;
&lt;/ul&gt;




</description>
      <category>webgl</category>
      <category>threejs</category>
      <category>ai</category>
      <category>showdev</category>
    </item>
    <item>
      <title>개발자 친구도 못 알아들어서, 제일 쉽게 다시 씁니다 — 스타일시드가 하는 일</title>
      <dc:creator>@kiwibreaksme</dc:creator>
      <pubDate>Thu, 16 Jul 2026 06:14:33 +0000</pubDate>
      <link>https://dev.to/kiwibreaksme/gaebalja-cingudo-mos-aladeuleoseo-jeil-swibge-dasi-sseubnida-seutailsideuga-haneun-il-36cg</link>
      <guid>https://dev.to/kiwibreaksme/gaebalja-cingudo-mos-aladeuleoseo-jeil-swibge-dasi-sseubnida-seutailsideuga-haneun-il-36cg</guid>
      <description>&lt;p&gt;어제 개발자 친구한테 카톡이 왔습니다.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"그 design md 잖아? 토스나 글로벌웹 토큰 추출한 거 샘플로 들어있고. 그럼 스타일시드를 쓰면 &lt;strong&gt;어떤 측면에서 일반 개발자들이 도움을 받음?&lt;/strong&gt;"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;10년 넘게 개발한 친구인데도 "디자인 룰 엔진"이라는 설명으로는 감이 안 온다는 겁니다. 제 잘못이죠. 그래서 그때 카톡으로 했던 설명을 그대로, 제일 쉬운 버전으로 다시 씁니다.&lt;/p&gt;

&lt;h2&gt;
  
  
  한 줄 요약
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;UI/UX 디자이너가 화면 만들 때 머릿속에서 자동으로 돌리는 판단들을, 룰 파일로 박아서, AI가 코딩할 때 강제로 지키게 하는 겁니다.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;토큰 샘플 모음이 아니라 판단 모음입니다. 차이가 뭐냐면 —&lt;/p&gt;

&lt;h2&gt;
  
  
  "판단"이 뭔데?
&lt;/h2&gt;

&lt;p&gt;디자이너는 화면을 만들 때 이런 걸 &lt;strong&gt;생각도 안 하고&lt;/strong&gt; 지킵니다. 몸에 박혀 있어서요:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. 검은색은 검은색을 안 쓴다.&lt;/strong&gt;&lt;br&gt;
&lt;code&gt;#000&lt;/code&gt; 순수 검정은 촌스럽습니다. 세련된 제품들은 전부 살짝 흐린 검정을 씁니다. 말로만 하면 안 믿으실 테니 실측 — 방금 toss.im의 CSS에서 색상값을 전부 뽑아봤습니다:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;#333d4b (18회) ← 본문
#191f28 (14회) ← 제일 어두운 텍스트
#4e5968 (10회)
#3182f6 (8회)  ← 토스 블루, 유일한 강조색
...
#000000 (0회)  ← 순검정은 단 한 번도 안 나옴
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;토스에서 제일 어두운 글자가 &lt;code&gt;#191F28&lt;/code&gt;입니다. 순검정이 아니라 살짝 푸른 기가 도는 짙은 회색이에요.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. 강조색은 딱 하나.&lt;/strong&gt;&lt;br&gt;
파란 버튼 옆에 보라색 링크, 그 옆에 초록 배지 — 개발자가 만들면 이렇게 됩니다. 디자이너는 강조색 하나만 정하고 나머지는 전부 회색조로 깝니다. 색이 하나여야 "그 색 = 눌러야 하는 것"이 되니까요.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. 폰트 크기는 아무 숫자나 쓰는 게 아니다.&lt;/strong&gt;&lt;br&gt;
15px, 17px, 19px... 필요할 때마다 즉흥으로 정하는 게 아니라, 정해진 표(예: 12·14·18·24·36·48)에서만 꺼내 씁니다. 그리고 크기마다 쓰임새가 정해져 있어요 — 48은 히어로 숫자, 18은 섹션 제목, 13은 날짜·캡션.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. 모서리를 각지게 했으면 끝까지 각지게.&lt;/strong&gt;&lt;br&gt;
각진 카드 안에 알약 버튼이 들어가는 순간 "두 제품 붙여놓은 느낌"이 납니다.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;5. 회색이 기본값, 색은 예외에만.&lt;/strong&gt;&lt;br&gt;
목록에서 정상 항목은 회색, 문제 있는 항목에만 색. 전부 색칠하면 아무것도 강조가 안 됩니다.&lt;/p&gt;

&lt;p&gt;이런 게 74개 있습니다. 하나하나는 사소해 보이는데, &lt;strong&gt;이걸 다 지킨 화면과 아닌 화면은 "전문가가 만든 것 vs AI가 뽑은 것"으로 갈립니다.&lt;/strong&gt; 커버 이미지가 그 차이예요 — 같은 제품, 같은 프롬프트, 룰만 다릅니다.&lt;/p&gt;
&lt;h2&gt;
  
  
  그런데 왜 "파일로 박아야" 하는가
&lt;/h2&gt;

&lt;p&gt;여기가 핵심입니다. 그냥 클로드로 코딩하면, &lt;strong&gt;컨텍스트가 끊기는 순간 통일성이 사라집니다.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;오늘 채팅에서 "강조색은 파랑 하나로 해줘"라고 말하면 그 화면은 그렇게 나옵니다. 그런데 다음 주에 페이지 하나 추가해달라고 하면? 그 대화 기억은 없으니 이번엔 보라색으로 뽑아옵니다. 폰트 크기도 저번엔 15px, 이번엔 16px. &lt;strong&gt;페이지를 추가할 때마다 조금씩 다른 앱이 됩니다.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;그래서 룰을 채팅이 아니라 &lt;strong&gt;프로젝트 안의 파일&lt;/strong&gt;(CLAUDE.md, AGENTS.md, .cursorrules)로 박습니다. AI 코딩 도구들은 이 파일을 매번 읽고 시작하거든요. 프로젝트별 결정(우리 키컬러는 #3182F6, 모서리는 부드럽게)은 별도 락 파일에 저장돼서, 다음 달에 페이지를 추가해도 같은 앱으로 나옵니다.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;일반 개발자가 받는 도움이 정확히 이겁니다: 디자이너를 고용하지 않고도, 디자이너가 옆에서 코드리뷰해주는 상태로 AI를 돌리는 것.&lt;/strong&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  이미 만든 사이트에도 씁니다
&lt;/h2&gt;

&lt;p&gt;새 화면을 만들 때만 쓰는 게 아닙니다. &lt;strong&gt;지금 있는 내 웹에 돌리면&lt;/strong&gt; — 얼마나 통일성 있게, 룰을 갖추고 디자인됐는지 &lt;strong&gt;점수로 뽑고&lt;/strong&gt;, 고칠 곳을 고쳐주고, 거기서 나온 결정들을 &lt;strong&gt;그 프로젝트의 룰로 만들어줍니다.&lt;/strong&gt; 한번 룰이 되면 이후 모든 수정이 그 룰을 따르고요.&lt;/p&gt;
&lt;h2&gt;
  
  
  하나 더: 룰만 주면 안 지킵니다
&lt;/h2&gt;

&lt;p&gt;몇 달 운영하면서 배운 건데, AI한테 룰 문서를 줘도 &lt;strong&gt;읽고, 끄덕이고, 첫 초안을 그대로 보여줍니다.&lt;/strong&gt; 사람이랑 똑같아요.&lt;/p&gt;

&lt;p&gt;그래서 채점기를 붙였습니다. 화면을 보여주기 전에 74개 룰 기준으로 0~100점을 매기고, 80점 미만이면 스스로 고친 다음에야 보여주게요. 참고로 이 채점기로 제 사이트를 채점했더니 58점이 나왔습니다. 고치는 과정까지 전부 공개해뒀어요: &lt;a href="https://styleseed-demo.vercel.app/scorecard" rel="noopener noreferrer"&gt;styleseed-demo.vercel.app/scorecard&lt;/a&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  30초 자가진단
&lt;/h2&gt;

&lt;p&gt;내 사이트가 "AI 티" 나는지 지금 확인할 수 있습니다. F12(개발자도구) 열고:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;CSS에서 &lt;code&gt;#000&lt;/code&gt; 검색 — 나오면 감점&lt;/li&gt;
&lt;li&gt;화면의 &lt;strong&gt;강조색 개수&lt;/strong&gt; 세기 — 2개 이상이면 감점&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;font-size&lt;/code&gt; 값 종류 세기 — 8종류 넘으면 감점&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;border-radius&lt;/code&gt; 값 종류 세기 — 제각각이면 감점&lt;/li&gt;
&lt;li&gt;목록/카드에서 &lt;strong&gt;색칠된 항목 비율&lt;/strong&gt; — 절반 넘게 색이 있으면 감점&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;3개 이상 걸리면 지금 "생성된 티"가 나고 있는 겁니다. (제 랜딩은 5개 중 4개 걸려서 58점이었습니다.)&lt;/p&gt;
&lt;h2&gt;
  
  
  써보기
&lt;/h2&gt;

&lt;p&gt;무료고 MIT 오픈소스입니다. Claude Code / Cursor / Codex에서:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx skills add bitjaru/styleseed
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;한 줄이면 붙습니다. 그다음 화면 만들 때 &lt;code&gt;/ss-build 대시보드&lt;/code&gt; 처럼 쓰면 위의 전 과정(락 → 빌드 → 채점 → 수정)이 자동으로 돕니다.&lt;/p&gt;

&lt;p&gt;→ &lt;a href="https://github.com/bitjaru/styleseed" rel="noopener noreferrer"&gt;github.com/bitjaru/styleseed&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;친구야, 이제 이해했냐?&lt;/p&gt;

</description>
      <category>design</category>
      <category>ai</category>
      <category>webdev</category>
      <category>vibecoding</category>
    </item>
    <item>
      <title>AI가 만든 UI, 보여주기 전에 채점시켜라 — 게이트를 만들고 내 랜딩부터 떨어진 이야기</title>
      <dc:creator>@kiwibreaksme</dc:creator>
      <pubDate>Wed, 15 Jul 2026 18:29:13 +0000</pubDate>
      <link>https://dev.to/kiwibreaksme/aiga-mandeun-ui-boyeojugi-jeone-caejeomsikyeora-geiteureul-mandeulgo-nae-raendingbuteo-ddeoleojin-iyagi-ea7</link>
      <guid>https://dev.to/kiwibreaksme/aiga-mandeun-ui-boyeojugi-jeone-caejeomsikyeora-geiteureul-mandeulgo-nae-raendingbuteo-ddeoleojin-iyagi-ea7</guid>
      <description>&lt;p&gt;&lt;em&gt;전편: &lt;a href="https://dev.to/kiwibreaksme/aiga-mandeun-uiga-eodinga-eosaeghan-iyu-geurigo-geugeol-gocineun-han-gaji-weoncig-5e4p"&gt;AI가 만든 UI가 어딘가 '어색한' 이유 — 그리고 그걸 고치는 한 가지 원칙&lt;/a&gt;에서 "축마다 값 하나"라는 정합성 원칙을 다뤘습니다. 이번 글은 그 다음 이야기 — *&lt;/em&gt;원칙만으론 부족했고, 강제 장치가 필요했다*&lt;em&gt;는 걸 배운 과정입니다.&lt;/em&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  룰을 줘도 안 지키더라
&lt;/h2&gt;

&lt;p&gt;전편을 쓰고 나서 저는 그 원칙들을 74개 룰로 정리해 오픈소스(&lt;a href="https://github.com/bitjaru/styleseed" rel="noopener noreferrer"&gt;StyleSeed&lt;/a&gt;)로 공개했습니다. Claude Code·Cursor·Codex가 UI를 만들기 전에 읽는 디자인 룰셋이죠.&lt;/p&gt;

&lt;p&gt;그런데 이상한 일이 생겼습니다. &lt;strong&gt;제가 만들면 잘 나오는데, 룰만 받아간 사람들 결과물은 여전히 "AI 티"가 났어요.&lt;/strong&gt; 같은 74개 룰인데요.&lt;/p&gt;

&lt;p&gt;몇 주를 파고든 결론: 격차는 룰이 아니라 &lt;strong&gt;루프&lt;/strong&gt;에 있었습니다. 저는 무의식적으로 이 순서를 돌고 있었어요.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;디자인 결정 고정(락) → 빌드 → 채점 → 80점 미만이면 수정 → 재채점 → 통과 후에만 확인
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;반면 룰만 받은 에이전트는 &lt;strong&gt;읽고, 끄덕이고, 첫 초안을 그대로 보여줍니다.&lt;/strong&gt; 사람으로 치면 코드리뷰 체크리스트를 벽에 붙여두고 리뷰 없이 머지하는 격이죠. 룰은 지식이고, 루프는 규율입니다. 지식은 전달됐는데 규율이 전달되지 않았던 겁니다.&lt;/p&gt;

&lt;h2&gt;
  
  
  그래서 루프를 강제했습니다
&lt;/h2&gt;

&lt;p&gt;v2.7부터 StyleSeed의 정문은 룰 문서가 아니라 &lt;strong&gt;강제 루프&lt;/strong&gt;입니다.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;/ss-build inventory dashboard
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;/ss-build&lt;/code&gt;는 이렇게 동작합니다:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;락 먼저&lt;/strong&gt; — 프로젝트에 &lt;code&gt;STYLESEED.md&lt;/code&gt;(키컬러·폰트·모서리·모션이 적힌 디자인 락 파일)가 없으면 &lt;strong&gt;UI 코드를 쓰기를 거부&lt;/strong&gt;하고 락부터 만듭니다. "AI가 만든 티"의 1등 원인이 매 프롬프트마다 결정이 초기화되는 드리프트거든요.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;빌드&lt;/strong&gt; — 락과 74개 룰 기준으로 만듭니다.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;게이트&lt;/strong&gt; — &lt;code&gt;/ss-score&lt;/code&gt;가 8개 카테고리(색 규율·위계·레이아웃·카드·상태·모션·정합성·독창성)로 0~100 채점하고, &lt;strong&gt;80점 미만이면 사용자에게 보여주지 않고&lt;/strong&gt; 감점 항목을 스스로 고친 뒤 재채점합니다.&lt;/li&gt;
&lt;li&gt;통과한 것만 최종 점수와 함께 보여줍니다.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;핵심 설계는 "80점은 천장이 아니라 바닥"이라는 것. 100점을 쫓느라 출시가 늦어지는 것도 실패 모드라서, 게이트는 명백히-제네릭한 초안이 사용자에게 도달하는 것만 막습니다.&lt;/p&gt;

&lt;h2&gt;
  
  
  코드는 통과했는데 픽셀이 어색하다면
&lt;/h2&gt;

&lt;p&gt;게이트를 몇 달 돌리면서 알게 된 것: &lt;strong&gt;최악의 "AI 티" 몇 개는 코드에 존재하지 않습니다.&lt;/strong&gt; 픽셀에만 존재해요.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;웹폰트가 조용히 실패해서 Times New Roman으로 렌더된 히어로&lt;/li&gt;
&lt;li&gt;코드상 문제없는데 화면 아래 1/3이 죽은 여백&lt;/li&gt;
&lt;li&gt;각각 정당한 두 색이 렌더되니 "강조색 2개"로 읽히는 화면&lt;/li&gt;
&lt;li&gt;포컬 포인트 없이 균일하게 나열된 카드 그리드&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;그래서 v2.10에 &lt;strong&gt;비주얼 게이트&lt;/strong&gt;(&lt;code&gt;/ss-verify&lt;/code&gt;)를 추가했습니다. 화면을 실제로 렌더 → 스크린샷 → 에이전트가 &lt;strong&gt;이미지를 직접 보고&lt;/strong&gt; 채점합니다. 코드 게이트(ss-score)와 픽셀 게이트(ss-verify), 두 단계 통과가 기본 루프가 됐습니다.&lt;/p&gt;

&lt;h2&gt;
  
  
  첫 희생자: 제 랜딩페이지 (58/100)
&lt;/h2&gt;

&lt;p&gt;게이트를 만들었으면 자기한테 먼저 돌려야죠. 게이트 도입 전에 만들었던 저희 랜딩 초기 버전을 넣었더니:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;카테고리&lt;/th&gt;
&lt;th&gt;점수&lt;/th&gt;
&lt;th&gt;잡힌 것&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;색 규율&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;7/16&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;언락 상태의 기본 인디고 + 그라디언트 히어로 텍스트&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;독창성&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;3/10&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;기능마다 반복되는 아이콘칩 클리셰, 목업 화면, 포컬 없음&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;정합성&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;5/12&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;반짝이 배지 + 그라디언트 헤드라인 + 플랫 버튼 — 언어 3개 혼재&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;위계·타이포&lt;/td&gt;
&lt;td&gt;12/16&lt;/td&gt;
&lt;td&gt;숫자/단위 2:1 미적용&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;상태·접근성&lt;/td&gt;
&lt;td&gt;12/18&lt;/td&gt;
&lt;td&gt;포커스 링 누락, 대비 미달&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;나머지 3개&lt;/td&gt;
&lt;td&gt;19/28&lt;/td&gt;
&lt;td&gt;오프그리드 간격, 1px 테두리 분리, 임의 페이드&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;합계&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;58/100 (F)&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;만든 사람 입장에서 꽤 굴욕적인데, 동시에 이 도구를 계속 만들 이유이기도 했습니다. &lt;strong&gt;게이트가 잡은 건 전부 기계적으로 고칠 수 있는 것들&lt;/strong&gt;이었거든요. 강조색 하나로 고정, 목업 대신 실제 제품 화면을 포컬로, 아이콘칩 카드를 넘버링 플로우로 — 58 → &lt;strong&gt;86&lt;/strong&gt;. 감점 내역 전체는 &lt;a href="https://styleseed-demo.vercel.app/scorecard" rel="noopener noreferrer"&gt;채점표 페이지&lt;/a&gt;에 공개해뒀습니다.&lt;/p&gt;

&lt;p&gt;커버 이미지가 그 비포/애프터입니다. 같은 제품, 같은 프롬프트 — 룰만 바꿨습니다.&lt;/p&gt;

&lt;h2&gt;
  
  
  Hacker News에서 또 한 번 (96→100)
&lt;/h2&gt;

&lt;p&gt;이 글을 쓰기 몇 시간 전, &lt;a href="https://news.ycombinator.com/item?id=48920888" rel="noopener noreferrer"&gt;Show HN&lt;/a&gt;에 올렸다가 이런 댓글을 받았습니다.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"Lighthouse 접근성 감사에서 떨어지는 요소가 꽤 있는데요."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;돌려보니 사실이었습니다. 96점, 실패 요소 19개 — 다크 섹션의 11px 라벨 대비 미달 16개, aria-label이 보이는 텍스트를 포함하지 않는 버튼 3개. &lt;strong&gt;접근성 채점 기준을 가진 도구의 홈페이지가, 최근 수정 후 게이트를 다시 안 돌려서&lt;/strong&gt; 드리프트한 겁니다.&lt;/p&gt;

&lt;p&gt;두 시간 안에 고쳐서 100점으로 만들고 답글로 보고했습니다. 교훈은 58점 때와 같습니다:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;게이트는 한 번 통과하는 시험이 아니라, 수정할 때마다 다시 돌리는 루프다.&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;자기 도구의 규칙을 자기가 어기는 건 부끄러운 일이 아니라 — 게이트가 없으면 누구나(만든 사람조차) 드리프트한다는 증거에 가깝습니다.&lt;/p&gt;

&lt;h2&gt;
  
  
  정합성 다음 단계: 프리셋
&lt;/h2&gt;

&lt;p&gt;전편의 반론 중 이런 게 있었습니다. &lt;em&gt;"정합성을 지켜도 여전히 제네릭해 보이면?"&lt;/em&gt; 맞는 말입니다. 일관된 화면도 얼마든지 밋밋할 수 있어요. 정합성은 필요조건이지 충분조건이 아닙니다.&lt;/p&gt;

&lt;p&gt;그래서 v2.8에 &lt;strong&gt;프리셋&lt;/strong&gt;(&lt;code&gt;/ss-restyle&lt;/code&gt;)을 넣었습니다 — swiss · editorial · technical · warm-dtc · minimal-mono · brutalist-lite. 각각이 모서리·타입·밀도·시그니처 무브까지 통째로 맞춰진 "하나의 좌표"라, 정합성을 깨지 않고 룩 전체를 갈아탑니다.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fc3k22qy1llbyh8fv7rhu.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fc3k22qy1llbyh8fv7rhu.png" alt="같은 제품을 6가지 프리셋으로" width="800" height="471"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;같은 제품이 구조부터 다른 6가지 화면이 됩니다. 각각 일관되고, 서로 다르고, 어느 것도 "기본값" 냄새가 나지 않는 것이 목표였습니다.&lt;/p&gt;

&lt;h2&gt;
  
  
  정리
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;룰만 주면 에이전트는 루프를 건너뛴다.&lt;/strong&gt; 지식이 아니라 규율을 전달해야 한다 — 게이트를 강제하는 커맨드로.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;코드 게이트 + 픽셀 게이트, 두 단계.&lt;/strong&gt; 최악의 티는 렌더된 화면에만 존재한다.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;게이트는 만든 사람부터 계속 통과해야 한다.&lt;/strong&gt; 내 랜딩 58점, HN 지적 후 접근성 96점 — 두 번 다 게이트를 건너뛴 내 잘못이었다.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;전부 오픈소스(MIT)입니다: &lt;strong&gt;&lt;a href="https://github.com/bitjaru/styleseed" rel="noopener noreferrer"&gt;github.com/bitjaru/styleseed&lt;/a&gt;&lt;/strong&gt; — &lt;code&gt;npx skills add bitjaru/styleseed&lt;/code&gt; 한 줄이면 Claude Code·Cursor·Codex에 붙습니다. 룰이 틀렸거나 너무 경직된 곳을 찾으시면 이슈로 알려주세요. 그게 이 룰셋이 자라는 방식입니다.&lt;/p&gt;

</description>
      <category>design</category>
      <category>ai</category>
      <category>webdev</category>
      <category>vibecoding</category>
    </item>
    <item>
      <title>I catalogued every tell that makes a UI look AI-generated. My own tool kept failing the test.</title>
      <dc:creator>@kiwibreaksme</dc:creator>
      <pubDate>Thu, 02 Jul 2026 05:23:12 +0000</pubDate>
      <link>https://dev.to/kiwibreaksme/i-catalogued-every-tell-that-makes-a-ui-look-ai-generated-my-own-tool-kept-failing-the-test-n52</link>
      <guid>https://dev.to/kiwibreaksme/i-catalogued-every-tell-that-makes-a-ui-look-ai-generated-my-own-tool-kept-failing-the-test-n52</guid>
      <description>&lt;p&gt;You can always tell when an AI built a UI. You can't always say &lt;em&gt;why&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;I maintain &lt;a href="https://github.com/bitjaru/styleseed" rel="noopener noreferrer"&gt;StyleSeed&lt;/a&gt; — an open-source set of design rules that Claude Code, Cursor, and Codex read automatically, so AI-built UI stops looking generated. For the past few months my actual job has been writing that "why" down: cataloguing every tell, turning each one into a rule an agent can obey, and then dogfooding the rules to see if they hold.&lt;/p&gt;

&lt;p&gt;They kept not holding — in ways that embarrassed me three separate times. That's what this post is about, because the failures turned out to be more useful than the rules.&lt;/p&gt;

&lt;h2&gt;
  
  
  The catalog: generation-one tells
&lt;/h2&gt;

&lt;p&gt;These are the ones everyone recognizes, even non-designers:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The default indigo.&lt;/strong&gt; &lt;code&gt;#4F46E5&lt;/code&gt; / &lt;code&gt;#5E6AD2&lt;/code&gt; — Tailwind's and every agent's comfort color. Zero brands chose it; every unguided agent ships it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The icon-in-a-pale-chip.&lt;/strong&gt; A generic Lucide line icon in an identical rounded square with a pastel background, repeated for &lt;em&gt;every&lt;/em&gt; feature card. One is fine. Seven is a signature.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Rainbow status lists.&lt;/strong&gt; Every row gets a differently-colored badge — including "normal," which should be grey. Color is supposed to mean &lt;em&gt;look here&lt;/em&gt;; paint every row and nothing means anything.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Emoji as UI icons.&lt;/strong&gt; 🚗🧺⭐ inject five uncontrolled hues each and break the single-accent rule instantly.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Centered everything + gradient headline + ✨ badge.&lt;/strong&gt; The stock hero.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;4 identical KPI cards. 3 identical pricing columns. 14px body on a 1440px screen. Pure #000 text.&lt;/strong&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Individually each is defensible. Together they compound into "an AI made this."&lt;/p&gt;

&lt;p&gt;So I wrote rules that ban each one — with the reasoning, so a model applies them instead of pattern-matching around them. Then I tested.&lt;/p&gt;

&lt;h2&gt;
  
  
  Failure #1: my own rules were &lt;em&gt;teaching&lt;/em&gt; the rainbow
&lt;/h2&gt;

&lt;p&gt;Dogfooding a mobile dashboard, the status list came out rainbow — the exact thing the project exists to prevent. I went digging expecting a model failure, and found this in my own DESIGN-LANGUAGE.md, rule 65:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Vary status states in lists to create &lt;strong&gt;visual interest through color diversity&lt;/strong&gt;.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;I was literally instructing agents to build rainbow lists. The agent had followed my rule perfectly. The rule was wrong.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2p29j10jvm8jagkjti7q.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2p29j10jvm8jagkjti7q.png" alt="Before: every status row a different color, " width="780" height="2198"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Rewrote rule 65 (status color = severity, never decoration), re-ran the same build: two grey "normal" rows, one amber, one red. Same agent, same prompt — only the rule changed.&lt;/p&gt;

&lt;h2&gt;
  
  
  Failure #2: my landing page scored 58/100 on my own quality gate
&lt;/h2&gt;

&lt;p&gt;StyleSeed ships a scored quality gate — the agent reviews its UI against the rules and fixes it to a floor of 80/100 &lt;em&gt;before showing you anything&lt;/em&gt;. At some point I did the obvious uncomfortable thing and ran the gate on &lt;strong&gt;StyleSeed's own landing page&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;58/100.&lt;/strong&gt; The page whose hero copy mocks "an icon in a pale chip on every card" had &lt;strong&gt;twelve icon chips&lt;/strong&gt; on it. Four accent colors. 11–13px text everywhere. A reviewer would have screenshotted it with the caption &lt;em&gt;"physician, heal thyself."&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;We rebuilt it to pass (one accent, chips gone, type floor met). But the lesson stuck: &lt;strong&gt;coherence rules don't enforce themselves — the gate has to run on everything, including your own marketing.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Failure #3: the escape hatches became clichés
&lt;/h2&gt;

&lt;p&gt;Here's the one I find genuinely interesting. Once you ban the generation-one tells, agents all flee through the same exits — and the exits are becoming the &lt;em&gt;next&lt;/em&gt; tells:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Ghost index numbers&lt;/strong&gt; (01 · 02 · 03) replacing every icon chip&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;UPPERCASE-overline + big-number cards&lt;/strong&gt; repeated identically for every KPI&lt;/li&gt;
&lt;li&gt;The &lt;strong&gt;text-left / visual-right hero with two pill CTAs&lt;/strong&gt; on every product site&lt;/li&gt;
&lt;li&gt;And the overshoot: escaping "generic" into &lt;strong&gt;dated&lt;/strong&gt; — a beige paper background + serif on everything, which stops reading "designed" and starts reading "government pamphlet." I shipped exactly this as a proud "after" example. It got rejected on sight — the generic &lt;em&gt;before&lt;/em&gt; honestly looked better at a glance.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;So the rulebook now tracks tells by generation: gen-1 (indigo, chips, rainbow) is banned outright; gen-2 has its own rule — &lt;em&gt;one signature treatment per project, varied section anatomy&lt;/em&gt;; gen-3 sets a modern floor — &lt;em&gt;white base, serif as seasoning, keep the air&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;This is the actual nature of the problem: it's not a checklist, it's an &lt;strong&gt;arms race against convergence&lt;/strong&gt;. Any escape route, once popular, becomes a tell.&lt;/p&gt;

&lt;h2&gt;
  
  
  Does it work?
&lt;/h2&gt;

&lt;p&gt;The cover image is the current proof: same product, same prompt, same model — only the rules changed. The before is an authentic agent-defaults build (competent, clean, and instantly recognizable as AI). The after chose its own accent, put the &lt;em&gt;real product&lt;/em&gt; in the hero, and varied every section's anatomy.&lt;/p&gt;

&lt;p&gt;I also ran a three-domain stress test with &lt;strong&gt;zero design direction&lt;/strong&gt; — a Korean mobile health app, a dark observability dashboard, a warm e-commerce landing. All three came out refined &lt;em&gt;and different from each other&lt;/em&gt;: calm teal + soft radii, signal-teal + IBM Plex + tonal dark, terracotta + a display grotesque. The thing I feared most — that the rules would just converge everything onto a new "StyleSeed look" — didn't happen.&lt;/p&gt;

&lt;h2&gt;
  
  
  What actually moves the needle
&lt;/h2&gt;

&lt;p&gt;After all the dogfooding, three mechanics matter more than any individual rule:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Constraints before code.&lt;/strong&gt; A &lt;code&gt;STYLESEED.md&lt;/code&gt; lock (mood, key color, font, surface) the agent re-reads every prompt. Unlocked = drift.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A scored gate with a floor.&lt;/strong&gt; "Review your work" is ignorable; "score it, fix to ≥80, then show me" is not.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ban lists with reasons.&lt;/strong&gt; Positive advice ("use nice spacing") does nothing. Named bans ("no &lt;code&gt;#4F46E5&lt;/code&gt;, no icon-chips, normal = grey") change output immediately.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;It's all MIT: &lt;a href="https://github.com/bitjaru/styleseed" rel="noopener noreferrer"&gt;github.com/bitjaru/styleseed&lt;/a&gt; — 74 rules + the gate + skins, readable by Claude Code / Cursor / Codex via one pasted sentence. Live demo at &lt;a href="https://styleseed-demo.vercel.app" rel="noopener noreferrer"&gt;styleseed-demo.vercel.app&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;What's the tell that instantly makes &lt;em&gt;you&lt;/em&gt; think "an AI made this"? I'm collecting them — the catalog only grows when someone names a new one.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>design</category>
      <category>webdev</category>
      <category>opensource</category>
    </item>
    <item>
      <title>AI가 만든 UI가 어딘가 '어색한' 이유 — 그리고 그걸 고치는 한 가지 원칙</title>
      <dc:creator>@kiwibreaksme</dc:creator>
      <pubDate>Tue, 16 Jun 2026 06:32:37 +0000</pubDate>
      <link>https://dev.to/kiwibreaksme/aiga-mandeun-uiga-eodinga-eosaeghan-iyu-geurigo-geugeol-gocineun-han-gaji-weoncig-5e4p</link>
      <guid>https://dev.to/kiwibreaksme/aiga-mandeun-uiga-eodinga-eosaeghan-iyu-geurigo-geugeol-gocineun-han-gaji-weoncig-5e4p</guid>
      <description>&lt;blockquote&gt;
&lt;p&gt;🇬🇧 English version: &lt;a href="https://dev.to/kiwibreaksme/why-ai-generated-uis-look-off-and-the-one-principle-that-fixes-it-4j20"&gt;Why AI-Generated UIs Look 'Off' — and the One Principle That Fixes It&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;🔄 &lt;strong&gt;후속편:&lt;/strong&gt; &lt;a href="https://dev.to/kiwibreaksme/aiga-mandeun-ui-boyeojugi-jeone-caejeomsikyeora-geiteureul-mandeulgo-nae-raendingbuteo-ddeoleojin-iyagi-ea7"&gt;AI가 만든 UI, 보여주기 전에 채점시켜라 — 게이트를 만들고 내 랜딩부터 떨어진 이야기&lt;/a&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;겪어보셨을 겁니다. AI한테 대시보드를 만들어 달라고 하면, 나오긴 나와요. 그런데… 묘하게 어색합니다. 컴포넌트 하나하나는 멀쩡해요. 버튼은 버튼답고, 카드는 카드답죠. 그런데 전체를 보면 &lt;em&gt;"생성된 티"&lt;/em&gt; 가 납니다. UI를 찍은 스톡 사진 같달까요.&lt;/p&gt;

&lt;p&gt;저도 한동안은 "컴포넌트를 더 잘 만들면" 또는 "프롬프트에 감각을 더 넣으면" 해결될 거라 생각했습니다. 아니더군요. 실제 디자인 문헌 — Refactoring UI, Material Design 3, Apple HIG, IBM Carbon, WCAG, Financial Times의 시각 어휘집 — 을 파고들수록 진짜 원인엔 이름이 있었습니다.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;정합성(coherence)의 부재.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;부품들이 서로 안 맞는 겁니다. 그리고 한번 보이기 시작하면, 다시는 안 보이게 할 수가 없어요.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;전에 &lt;a href="https://dev.to/kiwibreaksme/dijaineo-eobsi-baibeukodingeuro-toseugeub-ui-mandeuneun-beob-5c0g"&gt;디자이너 없이 바이브코딩으로 토스급 UI 만드는 법&lt;/a&gt;을 다뤘는데, 이번 글은 그 아래의 '왜' — "어색함"의 실제 메커니즘과 그걸 고치는 한 가지 원칙을 팝니다.&lt;/em&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  증상: 서로 안 맞는 부품들
&lt;/h2&gt;

&lt;p&gt;다음 UI엔 "틀린" 게 하나도 없습니다.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;모서리가 &lt;strong&gt;각진&lt;/strong&gt;(2px) 카드&lt;/li&gt;
&lt;li&gt;그 안의 &lt;strong&gt;알약 모양&lt;/strong&gt; 버튼&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;파란&lt;/strong&gt; 기본 버튼 옆의 &lt;strong&gt;보라&lt;/strong&gt; "업그레이드" 링크&lt;/li&gt;
&lt;li&gt;아이콘: Lucide 아웃라인 2개 + 채워진 이모지 1개&lt;/li&gt;
&lt;li&gt;좌상단에서 떨어진 그림자 하나, 정아래에서 떨어진 그림자 하나&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;하나씩 떼어 보면 전부 정당화됩니다. 그런데 합치면 "부품 조립" 신호가 떠요. 사람 디자이너라면 같은 화면에서 이 결정들을 동시에 내리지 않거든요. 디자이너는 모서리 스타일 &lt;em&gt;하나&lt;/em&gt;, 강조색 &lt;em&gt;하나&lt;/em&gt;, 아이콘 세트 &lt;em&gt;하나&lt;/em&gt;, 광원 &lt;em&gt;하나&lt;/em&gt; 를 고르고 — 그걸 어디서나 반복합니다.&lt;/p&gt;

&lt;p&gt;그 &lt;strong&gt;반복&lt;/strong&gt;이 우리가 "한 사람이 만든 느낌"으로 읽는 정체입니다. UX Collective의 Goran Paun은 이걸 시각적 정합성이라 부르고, Steve Schoger의 &lt;em&gt;Refactoring UI&lt;/em&gt; 전체가 어떤 의미에선 "불필요한 변주를 만들지 마라"는 책이죠. 원칙 자체는 오래됐습니다. 새로운 건 &lt;strong&gt;AI가 기본값으로 이걸 깨뜨린다&lt;/strong&gt;는 점이고, &lt;em&gt;왜&lt;/em&gt; 깨는지를 알면 어떻게 고칠지가 보입니다.&lt;/p&gt;




&lt;h2&gt;
  
  
  AI가 유독 이걸 못 하는 이유
&lt;/h2&gt;

&lt;p&gt;언어 모델은 UI를 &lt;strong&gt;국소적으로(locally)&lt;/strong&gt; 생성합니다. 카드 쓰고, 버튼 쓰고, 모달 쓰고 — 각각을 그럴듯한 새 조각으로 따로 만들어요. "40줄 위의 버튼이 알약 모양이었으니 이것도 알약이어야 한다"는 기억이 없습니다. 각 컴포넌트는 자기 학습 데이터의 평균값으로 회귀하는데, "버튼의 평균"과 "카드의 평균"은 애초에 서로 맞춰진 적이 없죠.&lt;/p&gt;

&lt;p&gt;사람은 이렇게 일하지 않습니다. 우리는 머릿속에 아주 작은 결정 집합 — &lt;em&gt;이 제품은 둥근 모서리, 파란 강조색, 8px 그리드&lt;/em&gt; — 을 들고 다니고, 새 요소는 전부 그걸 물려받습니다. 그 결정이 &lt;strong&gt;끈끈하게(sticky)&lt;/strong&gt; 따라붙어요.&lt;/p&gt;

&lt;p&gt;그래서 해법은 "모델을 더 감각 있게" 가 아닙니다. &lt;strong&gt;그 끈끈한 결정들을 적어두고, 모델에게 새로 만들지 말고 복사하라고 시키는 것&lt;/strong&gt; 입니다.&lt;/p&gt;




&lt;h2&gt;
  
  
  메타 법칙: 축마다 값 하나
&lt;/h2&gt;

&lt;p&gt;핵심은 한 문장입니다.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;디자인 축마다 값(또는 패밀리)을 &lt;strong&gt;하나만&lt;/strong&gt; 정하고, 토큰으로 박아서, &lt;strong&gt;어디서나&lt;/strong&gt; 적용한다.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;정합성은 "모든 화면이 똑같다"가 아닙니다. 같은 &lt;em&gt;결정&lt;/em&gt;이 반복되는 거예요. 설정 페이지와 대시보드는 전혀 다르게 생겨도 — 모서리 personality, 그림자 언어, 강조색, 간격 단위만 공유하면 — 한 제품처럼 느껴집니다.&lt;/p&gt;

&lt;p&gt;통일돼야 하는 축들:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;축&lt;/th&gt;
&lt;th&gt;하나만 정한다&lt;/th&gt;
&lt;th&gt;섞이면&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;모서리/라운드&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;각짐 0–4px · 부드러움 8–12px · 알약 9999px&lt;/td&gt;
&lt;td&gt;각진 다이얼로그 + 둥근 버튼 = "두 제품 붙인 느낌"&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;그림자&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;한 스케일, 한 광원(좌상단), 한 색조&lt;/td&gt;
&lt;td&gt;"태양이 둘인 장면"&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;강조색&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;강조는 1개 (+ 의미색 red/green/amber)&lt;/td&gt;
&lt;td&gt;무엇도 &lt;em&gt;그&lt;/em&gt; 액션으로 안 읽힘&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;간격 단위&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;8px 그리드 하나 (4px 하프스텝)&lt;/td&gt;
&lt;td&gt;7/13/19px 같은 오프그리드 = "엉성함"&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;아이콘&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;한 패밀리, 한 채움, 한 굵기&lt;/td&gt;
&lt;td&gt;세트 섞이면 "겉도는" 느낌&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;타입 스케일&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;모듈러 스케일 하나, 폰트 ≤2종&lt;/td&gt;
&lt;td&gt;임의 크기가 리듬을 깸&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;모션&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;지속시간 세트 하나 + 이징 패밀리 하나&lt;/td&gt;
&lt;td&gt;어떤 건 빠릿, 어떤 건 굼떠 = 다른 앱&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;컨트롤 높이&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;공유 높이 세트 (예: 40px)&lt;/td&gt;
&lt;td&gt;44px 인풋 옆 32px 버튼 = 베이스라인 깨짐&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;비전공자가 핵심을 가장 잘 찌른 표현:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"모서리를 각지게 했으면 다 각지게 가야지."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;정확합니다. 그리고 이건 위 표의 &lt;em&gt;모든&lt;/em&gt; 행에 그대로 적용돼요. &lt;strong&gt;축이 섞이는 건 스타일 선택이 아니라 lint 에러로 취급하세요.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;이 중 넷을 코드로 구체화해 봅시다.&lt;/p&gt;




&lt;h2&gt;
  
  
  1. 모서리 personality 하나 (그리고 중첩 법칙)
&lt;/h2&gt;

&lt;p&gt;먼저, 라운드는 &lt;strong&gt;토큰&lt;/strong&gt;이지 매직 넘버가 아닙니다. 작은 스케일을 정의하고 어디서나 참조하세요.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight css"&gt;&lt;code&gt;&lt;span class="nd"&gt;:root&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="py"&gt;--radius-sm&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="m"&gt;8px&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;   &lt;span class="c"&gt;/* 인풋, 버튼  */&lt;/span&gt;
  &lt;span class="py"&gt;--radius-md&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="m"&gt;12px&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;  &lt;span class="c"&gt;/* 카드, 메뉴  */&lt;/span&gt;
  &lt;span class="py"&gt;--radius-lg&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="m"&gt;16px&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;  &lt;span class="c"&gt;/* 모달, 시트  */&lt;/span&gt;
  &lt;span class="py"&gt;--radius-full&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="m"&gt;9999px&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;그다음 &lt;strong&gt;personality 하나&lt;/strong&gt; — 전부 각지게, 전부 부드럽게, 전부 알약 — 를 골라 카드/버튼/인풋/모달/이미지에 적용합니다. 모델이 &lt;code&gt;rounded-none&lt;/code&gt; 패널에 &lt;code&gt;rounded-full&lt;/code&gt; 버튼을 멋대로 섞게 두지 마세요.&lt;/p&gt;

&lt;p&gt;많이들 놓치는 두 번째 규칙: &lt;strong&gt;중첩된 라운드는 중심을 공유해야 합니다.&lt;/strong&gt; 둥근 컨테이너 안에 패딩을 두고 둥근 요소가 들어갈 때, 안쪽 반지름은 &lt;em&gt;더 작아야&lt;/em&gt; 해요.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;안쪽 반지름 = 바깥 반지름 − 패딩
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight css"&gt;&lt;code&gt;&lt;span class="nc"&gt;.card&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="py"&gt;--pad&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="m"&gt;16px&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="nl"&gt;border-radius&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;var&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;--radius-lg&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;          &lt;span class="c"&gt;/* 16px */&lt;/span&gt;
  &lt;span class="nl"&gt;padding&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;var&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;--pad&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="nc"&gt;.card&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="nc"&gt;.thumbnail&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="c"&gt;/* 16 − 16 = 0; 음수가 안 되게 clamp */&lt;/span&gt;
  &lt;span class="nl"&gt;border-radius&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;max&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="m"&gt;0px&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;calc&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;var&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;--radius-lg&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="n"&gt;-&lt;/span&gt; &lt;span class="n"&gt;var&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;--pad&lt;/span&gt;&lt;span class="p"&gt;)));&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;안쪽 요소에 바깥과 &lt;em&gt;같은&lt;/em&gt; 라운드를 주면, 코너가 컨테이너의 호를 비집고 튀어나옵니다. Apple의 새 "Liquid Glass"가 바로 이 동심원 코너 수학을 공식화했고, Cloud Four에 정석 글이 있습니다.&lt;/p&gt;




&lt;h2&gt;
  
  
  2. 단일 광원의 레이어드 그림자
&lt;/h2&gt;

&lt;p&gt;깊이를 가짜로 보이게 하는 가장 빠른 방법이 단일 하드 &lt;code&gt;box-shadow&lt;/code&gt;입니다. 진짜 그림자는 그라데이션(반음영)이라, 낮은 불투명도 레이어를 여러 겹 쌓고 — 결정적으로 — 검정 대신 &lt;strong&gt;표면 색조 쪽으로 살짝 틴트&lt;/strong&gt;하세요. 순수 검정은 탁해집니다.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight css"&gt;&lt;code&gt;&lt;span class="c"&gt;/* 카드 — 페이지에 가깝게 */&lt;/span&gt;
&lt;span class="nt"&gt;--shadow-md&lt;/span&gt;&lt;span class="o"&gt;:&lt;/span&gt;
  &lt;span class="err"&gt;0&lt;/span&gt; &lt;span class="err"&gt;1&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt; &lt;span class="err"&gt;2&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt;  &lt;span class="nt"&gt;hsl&lt;/span&gt;&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="err"&gt;220&lt;/span&gt; &lt;span class="err"&gt;40&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="err"&gt;20&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="err"&gt;0&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="err"&gt;08&lt;/span&gt;&lt;span class="o"&gt;),&lt;/span&gt;
  &lt;span class="err"&gt;0&lt;/span&gt; &lt;span class="err"&gt;2&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt; &lt;span class="err"&gt;4&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt;  &lt;span class="nt"&gt;hsl&lt;/span&gt;&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="err"&gt;220&lt;/span&gt; &lt;span class="err"&gt;40&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="err"&gt;20&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="err"&gt;0&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="err"&gt;08&lt;/span&gt;&lt;span class="o"&gt;),&lt;/span&gt;
  &lt;span class="err"&gt;0&lt;/span&gt; &lt;span class="err"&gt;4&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt; &lt;span class="err"&gt;8&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt;  &lt;span class="nt"&gt;hsl&lt;/span&gt;&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="err"&gt;220&lt;/span&gt; &lt;span class="err"&gt;40&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="err"&gt;20&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="err"&gt;0&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="err"&gt;08&lt;/span&gt;&lt;span class="o"&gt;);&lt;/span&gt;

&lt;span class="c"&gt;/* 모달 — 사용자 쪽으로 떠오르게 */&lt;/span&gt;
&lt;span class="nt"&gt;--shadow-xl&lt;/span&gt;&lt;span class="o"&gt;:&lt;/span&gt;
  &lt;span class="err"&gt;0&lt;/span&gt; &lt;span class="err"&gt;2&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt; &lt;span class="err"&gt;4&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt;   &lt;span class="nt"&gt;hsl&lt;/span&gt;&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="err"&gt;220&lt;/span&gt; &lt;span class="err"&gt;40&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="err"&gt;20&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="err"&gt;0&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="err"&gt;06&lt;/span&gt;&lt;span class="o"&gt;),&lt;/span&gt;
  &lt;span class="err"&gt;0&lt;/span&gt; &lt;span class="err"&gt;8&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt; &lt;span class="err"&gt;16&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt;  &lt;span class="nt"&gt;hsl&lt;/span&gt;&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="err"&gt;220&lt;/span&gt; &lt;span class="err"&gt;40&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="err"&gt;20&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="err"&gt;0&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="err"&gt;06&lt;/span&gt;&lt;span class="o"&gt;),&lt;/span&gt;
  &lt;span class="err"&gt;0&lt;/span&gt; &lt;span class="err"&gt;16&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt; &lt;span class="err"&gt;32&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt; &lt;span class="nt"&gt;hsl&lt;/span&gt;&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="err"&gt;220&lt;/span&gt; &lt;span class="err"&gt;40&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="err"&gt;20&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="err"&gt;0&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="err"&gt;06&lt;/span&gt;&lt;span class="o"&gt;),&lt;/span&gt;
  &lt;span class="err"&gt;0&lt;/span&gt; &lt;span class="err"&gt;32&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt; &lt;span class="err"&gt;64&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt; &lt;span class="nt"&gt;hsl&lt;/span&gt;&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="err"&gt;220&lt;/span&gt; &lt;span class="err"&gt;40&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="err"&gt;20&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="err"&gt;0&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="err"&gt;06&lt;/span&gt;&lt;span class="o"&gt;);&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;정합성 규칙: &lt;strong&gt;페이지 전체가 한 광원&lt;/strong&gt; (관례상 위 + 약간 왼쪽), 수직 오프셋은 수평의 약 2배. 고도가 올라가면 오프셋·블러는 커지고 불투명도는 낮아집니다. 그러면 카드와 모달이 &lt;em&gt;같은&lt;/em&gt; 빛이 다른 높이의 물체에 닿은 걸로 읽혀요 — 별개의 두 효과가 아니라. (Josh Comeau, Tobias Ahlin 글이 좋습니다.)&lt;/p&gt;

&lt;p&gt;다크 모드에선 그림자가 거의 안 보이므로 &lt;strong&gt;톤 기반 고도&lt;/strong&gt;로 전환하세요 — 높을수록 밝은 표면. 베이스는 절대 순수 &lt;code&gt;#000&lt;/code&gt; 금지(Material은 &lt;code&gt;#121212&lt;/code&gt; + 레벨별 화이트 오버레이 권장).&lt;/p&gt;




&lt;h2&gt;
  
  
  3. 강조색 하나, 나머지는 그레이스케일
&lt;/h2&gt;

&lt;p&gt;"AI 무지개" UI는 대개 강조가 필요할 때마다 새 색을 꺼내 써서 생깁니다. 규율:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;상호작용 강조는 &lt;strong&gt;하나&lt;/strong&gt;의 강조색.&lt;/li&gt;
&lt;li&gt;텍스트·표면·테두리는 &lt;strong&gt;틴트된 그레이 램프&lt;/strong&gt; (순수 회색 말고 — 브랜드 색조 쪽으로 5–15% 밀기).&lt;/li&gt;
&lt;li&gt;의미색은 &lt;strong&gt;딱 넷&lt;/strong&gt; — success / warning / error / info — 이고 &lt;em&gt;오직&lt;/em&gt; 의미로만, 장식으로는 절대 안 씀.
&lt;/li&gt;
&lt;/ul&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight css"&gt;&lt;code&gt;&lt;span class="nd"&gt;:root&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="py"&gt;--accent&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="m"&gt;#5b5bd6&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="c"&gt;/* 채도 0 회색이 아니라 강조색 쪽으로 살짝 틴트 */&lt;/span&gt;
  &lt;span class="py"&gt;--grey-50&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;  &lt;span class="n"&gt;hsl&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="m"&gt;240&lt;/span&gt; &lt;span class="m"&gt;20%&lt;/span&gt; &lt;span class="m"&gt;98%&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
  &lt;span class="py"&gt;--grey-200&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;hsl&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="m"&gt;240&lt;/span&gt; &lt;span class="m"&gt;14%&lt;/span&gt; &lt;span class="m"&gt;90%&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
  &lt;span class="py"&gt;--grey-500&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;hsl&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="m"&gt;240&lt;/span&gt; &lt;span class="m"&gt;8%&lt;/span&gt;  &lt;span class="m"&gt;55%&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
  &lt;span class="py"&gt;--grey-900&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;hsl&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="m"&gt;240&lt;/span&gt; &lt;span class="m"&gt;12%&lt;/span&gt; &lt;span class="m"&gt;14%&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;  &lt;span class="c"&gt;/* 본문 텍스트 — #000 아님 */&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;그리고 타협 불가 접근성 하한선(WCAG 2.2 AA): 본문 텍스트 &lt;strong&gt;4.5:1&lt;/strong&gt;, 큰 텍스트·UI 요소 &lt;strong&gt;3:1&lt;/strong&gt;. 정보를 색으로만 전달 금지 — 아이콘·텍스트·형태와 함께. 유효성 실패 필드에 빨간 테두리만으론 부족하고, 아이콘과 메시지를 같이 넣으세요.&lt;/p&gt;




&lt;h2&gt;
  
  
  4. 간격 그리드 하나, 그룹핑은 근접성으로
&lt;/h2&gt;

&lt;p&gt;모든 margin·padding·gap을 한 스케일 — &lt;code&gt;4, 8, 12, 16, 24, 32, 48, 64&lt;/code&gt; — 에 스냅하고, &lt;strong&gt;근접성&lt;/strong&gt;이 의미를 나르게 하세요. Gestalt(그리고 Refactoring UI의 "모호한 간격을 피하라")의 규칙: 그룹 &lt;em&gt;바깥&lt;/em&gt; 공간이 그룹 &lt;em&gt;안&lt;/em&gt; 공간보다 확실히 커야 합니다.&lt;/p&gt;

&lt;p&gt;폼의 구체적 사다리:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;라벨 → 인풋        : 4–8px
인풋 → 인풋         : 12–16px
그룹 → 그룹 (섹션)  : 24–32px
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;전부 균일하게 띄우면 눈이 무엇이 한 덩어리인지 알 수 없습니다. 단계마다 두 배로 벌어지는 게, 라벨이 &lt;em&gt;자기&lt;/em&gt; 필드에 붙고 섹션이 다음 섹션과 분리되게 만드는 핵심이에요.&lt;/p&gt;




&lt;h2&gt;
  
  
  기계가 스스로 점검하게
&lt;/h2&gt;

&lt;p&gt;결정을 적어두는 게 절반이고, 나머지 절반은 위반을 잡는 겁니다. 여기서 AI가 실제로 쓸모 있어요 — 채점 기준만 쥐여주면.&lt;/p&gt;

&lt;p&gt;저는 위 내용을 Claude Code·Cursor가 자동으로 읽는 오픈소스 디자인 엔진(&lt;a href="https://github.com/bitjaru/styleseed" rel="noopener noreferrer"&gt;StyleSeed&lt;/a&gt;)에 넣지만, 아이디어 자체는 도구 무관입니다. 가장 쓸모 있던 조각은 &lt;strong&gt;정합성 채점기&lt;/strong&gt; — 파일을 점수화하고 &lt;em&gt;섞인 축&lt;/em&gt;을 실제 감점으로 잡는 검사였습니다.&lt;/p&gt;

&lt;p&gt;출력 예시(축약):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight markdown"&gt;&lt;code&gt;&lt;span class="gu"&gt;## Design Score: 70 / 100   (Dashboard.tsx)&lt;/span&gt;

Color discipline   13/18  ▓▓▓░  #000 제목; 강조색 3개
Cards &amp;amp; elevation   8/12  ▓▓░░  1px 테두리가 분리를 담당
Coherence           6/12  ▓▓░░  각진 카드(l.22) + 알약 버튼(l.48); 강조 3개

&lt;span class="gu"&gt;### 먼저 고칠 것 (점수 이득 큰 순)&lt;/span&gt;
&lt;span class="p"&gt;1.&lt;/span&gt; 라운드 통일(부드러움 8–12px) + 강조색 1개로  → +9
&lt;span class="p"&gt;2.&lt;/span&gt; 주문 리스트에 empty + loading 상태 추가       → +7
&lt;span class="p"&gt;3.&lt;/span&gt; 1px 테두리 빼고 톤 + ≤8% 그림자               → +4
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;"Coherence" 항목이 "AI 생성 티"를 가장 잘 예측합니다. 컴포넌트별 예쁨이 아니라 &lt;em&gt;전역 일관성&lt;/em&gt;을 재기 때문이에요. 컴포넌트는 예뻐도 틀릴 수 있습니다 — 이웃과 안 맞으면.&lt;/p&gt;




&lt;h2&gt;
  
  
  결론
&lt;/h2&gt;

&lt;p&gt;하나만 기억한다면:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;예쁜 부품이 예쁜 UI를 만들지 않습니다. &lt;strong&gt;서로 맞는&lt;/strong&gt; 부품이 만듭니다.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;축마다 값 하나 — 라운드, 그림자, 강조색, 간격, 아이콘, 타입, 모션 — 를 골라 토큰으로 적어두고, 새 요소가 매번 새로 고르는 대신 그걸 물려받게 하세요. 그 한 가지 규율이 "로봇이 만들었네"와 "디자이너가 만들었네"를 가르고, AI가 만든 인터페이스에 할 수 있는 가장 싸고 효과 큰 일입니다.&lt;/p&gt;

&lt;p&gt;모서리가 각지다고요? 그럼 전부 각지게. 끝까지.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;숫자까지 들어간 풀버전 — 간격 수치, 앱 종류별 타입 레시피, 레이어드 그림자·중첩 라운드 레시피, 전부 Refactoring UI / Material 3 / Apple HIG / WCAG 근거 — 은 오픈소스(MIT)로 여기 있습니다: &lt;a href="https://github.com/bitjaru/styleseed" rel="noopener noreferrer"&gt;github.com/bitjaru/styleseed&lt;/a&gt;. ⭐ 하나가 더 많은 개발자(와 AI 도구)에게 닿게 합니다.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;🇬🇧 The English version of this post is &lt;a href="https://dev.to/kiwibreaksme/why-ai-generated-uis-look-off-and-the-one-principle-that-fixes-it-4j20"&gt;here&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>design</category>
      <category>css</category>
      <category>webdev</category>
      <category>ai</category>
    </item>
    <item>
      <title>Why AI-Generated UIs Look 'Off' — and the One Principle That Fixes It</title>
      <dc:creator>@kiwibreaksme</dc:creator>
      <pubDate>Tue, 16 Jun 2026 04:33:17 +0000</pubDate>
      <link>https://dev.to/kiwibreaksme/why-ai-generated-uis-look-off-and-the-one-principle-that-fixes-it-4j20</link>
      <guid>https://dev.to/kiwibreaksme/why-ai-generated-uis-look-off-and-the-one-principle-that-fixes-it-4j20</guid>
      <description>&lt;p&gt;🇰🇷 한국어 버전: &lt;a href="https://dev.to/kiwibreaksme/aiga-mandeun-uiga-eodinga-eosaeghan-iyu-geurigo-geugeol-gocineun-han-gaji-weoncig-5e4p"&gt;AI가 만든 UI가 어딘가 '어색한' 이유&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;You've seen it. You ask an AI to build a dashboard, and it returns something that's… fine? Every component is individually competent. The button is a button. The card is a card. And yet the whole thing reads as &lt;em&gt;generated&lt;/em&gt; — like a stock photo of a UI.&lt;/p&gt;

&lt;p&gt;For a long time I assumed the fix was "better components" or "more taste in the prompt." It isn't. After digging through the actual design literature — Refactoring UI, Material Design 3, Apple's HIG, IBM Carbon, WCAG, the Financial Times' visual vocabulary — the real culprit has a name:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Incoherence.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;The parts don't agree with each other. And once you can see it, you can't unsee it.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;I've written before about &lt;a href="https://dev.to/kiwibreaksme/your-vibe-coded-app-looks-ugly-heres-what-i-did-about-it-2nb4"&gt;why your vibe-coded app looks ugly and what I did about it&lt;/a&gt;. This post goes underneath that — the actual mechanism behind the "off" feeling, and the one principle that fixes it.&lt;/em&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  The tell: parts that don't agree
&lt;/h2&gt;

&lt;p&gt;Here's a UI with nothing "wrong" in it:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;A card with &lt;strong&gt;sharp&lt;/strong&gt; 2px corners&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Pill-shaped&lt;/strong&gt; buttons inside it&lt;/li&gt;
&lt;li&gt;A &lt;strong&gt;blue&lt;/strong&gt; primary action next to a &lt;strong&gt;purple&lt;/strong&gt; "upgrade" link&lt;/li&gt;
&lt;li&gt;Icons: two from Lucide (outline), one filled emoji&lt;/li&gt;
&lt;li&gt;One shadow lit from the top-left, another from straight below&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Every one of those choices is defensible in isolation. Together, they signal "assembled from parts," because a human designer would never make all of them on the same screen. A designer picks &lt;em&gt;one&lt;/em&gt; corner style, &lt;em&gt;one&lt;/em&gt; accent, &lt;em&gt;one&lt;/em&gt; icon set, &lt;em&gt;one&lt;/em&gt; light source — and repeats those decisions everywhere.&lt;/p&gt;

&lt;p&gt;That repetition is what we read as "designed by one mind." Goran Paun (UX Collective) calls it visual coherence; Steve Schoger's whole &lt;em&gt;Refactoring UI&lt;/em&gt; is, in a sense, a book about not introducing gratuitous variation. The principle is old. What's new is that &lt;strong&gt;AI breaks it by default&lt;/strong&gt;, and understanding &lt;em&gt;why&lt;/em&gt; tells you how to fix it.&lt;/p&gt;




&lt;h2&gt;
  
  
  Why AI specifically struggles with this
&lt;/h2&gt;

&lt;p&gt;A language model generates UI &lt;strong&gt;locally&lt;/strong&gt;. It writes the card, then the button, then the modal — each as a fresh, plausible snippet. It has no running memory that "the buttons 40 lines up were pill-shaped, so these must be too." Each component regresses to the mean of its training data, and the mean of "a button" and the mean of "a card" were never coordinated.&lt;/p&gt;

&lt;p&gt;Humans don't work this way. We carry a tiny set of decisions in our head — &lt;em&gt;this product is soft-cornered, blue-accented, 8px-grid&lt;/em&gt; — and every new element inherits them. The decisions are sticky.&lt;/p&gt;

&lt;p&gt;So the fix isn't "make the model more tasteful." It's: &lt;strong&gt;write the sticky decisions down, and tell the model to copy them instead of inventing new ones.&lt;/strong&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  The meta-law: one choice per axis
&lt;/h2&gt;

&lt;p&gt;Here's the whole idea in one sentence:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;For each design axis, pick exactly &lt;strong&gt;one&lt;/strong&gt; value or family, encode it as a token, and apply it &lt;strong&gt;everywhere&lt;/strong&gt;.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Coherence is not "every screen is identical." It's that the same &lt;em&gt;decisions&lt;/em&gt; recur. A settings page and a dashboard can look very different and still feel like one product — as long as they share a radius personality, a shadow language, an accent, and a spacing unit.&lt;/p&gt;

&lt;p&gt;These are the axes that have to stay unified:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Axis&lt;/th&gt;
&lt;th&gt;Pick ONE, system-wide&lt;/th&gt;
&lt;th&gt;Failure mode when mixed&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Corner / radius&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;sharp 0–4px · soft 8–12px · pill 9999px&lt;/td&gt;
&lt;td&gt;sharp dialog + rounded buttons = "two products glued together"&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Shadow&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;one scale, one light source (above-left), one tint&lt;/td&gt;
&lt;td&gt;"a scene with two suns"&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Accent color&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;one accent for emphasis (+ semantic red/green/amber)&lt;/td&gt;
&lt;td&gt;nothing reads as &lt;em&gt;the&lt;/em&gt; action&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Spacing unit&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;one 8px grid (4px half-step)&lt;/td&gt;
&lt;td&gt;off-grid 7/13/19px reads as "sloppy"&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Icon style&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;one family, one fill mode, one stroke weight&lt;/td&gt;
&lt;td&gt;mixing sets looks "out of place"&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Type scale&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;one modular scale, ≤2 families&lt;/td&gt;
&lt;td&gt;arbitrary sizes destroy rhythm&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Motion&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;one duration set + one easing family&lt;/td&gt;
&lt;td&gt;some snappy, some sluggish = different apps&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Control height&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;shared height set (e.g. 40px)&lt;/td&gt;
&lt;td&gt;a 44px input beside a 32px button breaks the baseline&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The single best instinct I've heard a non-designer voice it as:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"If the corners are sharp, everything should be sharp."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Exactly right — and it generalizes to &lt;em&gt;every&lt;/em&gt; row in that table. &lt;strong&gt;Treat a mixed axis as a lint error, not a style choice.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F3l2zltx9sf7kih8dxwe0.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F3l2zltx9sf7kih8dxwe0.png" alt="Every decision in a StyleSeed UI tied to the rule behind it — numbers 2:1, one accent, content in cards, one corner radius." width="799" height="453"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Let's make four of these concrete.&lt;/p&gt;




&lt;h2&gt;
  
  
  1. Pick one corner personality (and respect the nesting law)
&lt;/h2&gt;

&lt;p&gt;First, radius is a &lt;strong&gt;token&lt;/strong&gt;, never a magic number. Define a small scale and reference it everywhere:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight css"&gt;&lt;code&gt;&lt;span class="nd"&gt;:root&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="py"&gt;--radius-sm&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="m"&gt;8px&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;   &lt;span class="c"&gt;/* inputs, buttons   */&lt;/span&gt;
  &lt;span class="py"&gt;--radius-md&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="m"&gt;12px&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;  &lt;span class="c"&gt;/* cards, menus      */&lt;/span&gt;
  &lt;span class="py"&gt;--radius-lg&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="m"&gt;16px&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;  &lt;span class="c"&gt;/* modals, sheets    */&lt;/span&gt;
  &lt;span class="py"&gt;--radius-full&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="m"&gt;9999px&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then pick &lt;strong&gt;one personality&lt;/strong&gt; — all sharp, all soft, or all pill — and apply it across card / button / input / modal / image. Don't let the model freestyle a &lt;code&gt;rounded-none&lt;/code&gt; panel with &lt;code&gt;rounded-full&lt;/code&gt; buttons.&lt;/p&gt;

&lt;p&gt;There's a subtle second rule most people miss: &lt;strong&gt;nested radii must share a center.&lt;/strong&gt; When a rounded element sits inside a rounded container with padding, the inner radius should be &lt;em&gt;smaller&lt;/em&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;inner radius = outer radius − padding
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight css"&gt;&lt;code&gt;&lt;span class="nc"&gt;.card&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="py"&gt;--pad&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="m"&gt;16px&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="nl"&gt;border-radius&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;var&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;--radius-lg&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;          &lt;span class="c"&gt;/* 16px */&lt;/span&gt;
  &lt;span class="nl"&gt;padding&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;var&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;--pad&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="nc"&gt;.card&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="nc"&gt;.thumbnail&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="c"&gt;/* 16 − 16 = 0; clamp so it never goes negative */&lt;/span&gt;
  &lt;span class="nl"&gt;border-radius&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;max&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="m"&gt;0px&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;calc&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;var&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;--radius-lg&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="n"&gt;-&lt;/span&gt; &lt;span class="n"&gt;var&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;--pad&lt;/span&gt;&lt;span class="p"&gt;)));&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If you give the inner element the &lt;em&gt;same&lt;/em&gt; radius as the outer one, its corner visibly bulges past the container's arc. Apple's new "Liquid Glass" system formalizes exactly this concentric-corner math; Cloud Four has the canonical write-up.&lt;/p&gt;




&lt;h2&gt;
  
  
  2. Layered shadows from a single light source
&lt;/h2&gt;

&lt;p&gt;The fastest way to make depth look fake is one hard &lt;code&gt;box-shadow&lt;/code&gt;. Real shadows are a gradient (a penumbra), so stack several low-opacity layers, and — critically — &lt;strong&gt;tint them toward the surface hue&lt;/strong&gt; instead of using pure black, which goes muddy:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight css"&gt;&lt;code&gt;&lt;span class="c"&gt;/* card — sits near the page */&lt;/span&gt;
&lt;span class="nt"&gt;--shadow-md&lt;/span&gt;&lt;span class="o"&gt;:&lt;/span&gt;
  &lt;span class="err"&gt;0&lt;/span&gt; &lt;span class="err"&gt;1&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt; &lt;span class="err"&gt;2&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt;  &lt;span class="nt"&gt;hsl&lt;/span&gt;&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="err"&gt;220&lt;/span&gt; &lt;span class="err"&gt;40&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="err"&gt;20&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="err"&gt;0&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="err"&gt;08&lt;/span&gt;&lt;span class="o"&gt;),&lt;/span&gt;
  &lt;span class="err"&gt;0&lt;/span&gt; &lt;span class="err"&gt;2&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt; &lt;span class="err"&gt;4&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt;  &lt;span class="nt"&gt;hsl&lt;/span&gt;&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="err"&gt;220&lt;/span&gt; &lt;span class="err"&gt;40&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="err"&gt;20&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="err"&gt;0&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="err"&gt;08&lt;/span&gt;&lt;span class="o"&gt;),&lt;/span&gt;
  &lt;span class="err"&gt;0&lt;/span&gt; &lt;span class="err"&gt;4&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt; &lt;span class="err"&gt;8&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt;  &lt;span class="nt"&gt;hsl&lt;/span&gt;&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="err"&gt;220&lt;/span&gt; &lt;span class="err"&gt;40&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="err"&gt;20&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="err"&gt;0&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="err"&gt;08&lt;/span&gt;&lt;span class="o"&gt;);&lt;/span&gt;

&lt;span class="c"&gt;/* modal — floats toward the user */&lt;/span&gt;
&lt;span class="nt"&gt;--shadow-xl&lt;/span&gt;&lt;span class="o"&gt;:&lt;/span&gt;
  &lt;span class="err"&gt;0&lt;/span&gt; &lt;span class="err"&gt;2&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt; &lt;span class="err"&gt;4&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt;   &lt;span class="nt"&gt;hsl&lt;/span&gt;&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="err"&gt;220&lt;/span&gt; &lt;span class="err"&gt;40&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="err"&gt;20&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="err"&gt;0&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="err"&gt;06&lt;/span&gt;&lt;span class="o"&gt;),&lt;/span&gt;
  &lt;span class="err"&gt;0&lt;/span&gt; &lt;span class="err"&gt;8&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt; &lt;span class="err"&gt;16&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt;  &lt;span class="nt"&gt;hsl&lt;/span&gt;&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="err"&gt;220&lt;/span&gt; &lt;span class="err"&gt;40&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="err"&gt;20&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="err"&gt;0&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="err"&gt;06&lt;/span&gt;&lt;span class="o"&gt;),&lt;/span&gt;
  &lt;span class="err"&gt;0&lt;/span&gt; &lt;span class="err"&gt;16&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt; &lt;span class="err"&gt;32&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt; &lt;span class="nt"&gt;hsl&lt;/span&gt;&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="err"&gt;220&lt;/span&gt; &lt;span class="err"&gt;40&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="err"&gt;20&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="err"&gt;0&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="err"&gt;06&lt;/span&gt;&lt;span class="o"&gt;),&lt;/span&gt;
  &lt;span class="err"&gt;0&lt;/span&gt; &lt;span class="err"&gt;32&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt; &lt;span class="err"&gt;64&lt;/span&gt;&lt;span class="nt"&gt;px&lt;/span&gt; &lt;span class="nt"&gt;hsl&lt;/span&gt;&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="err"&gt;220&lt;/span&gt; &lt;span class="err"&gt;40&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="err"&gt;20&lt;/span&gt;&lt;span class="o"&gt;%&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="err"&gt;0&lt;/span&gt;&lt;span class="o"&gt;.&lt;/span&gt;&lt;span class="err"&gt;06&lt;/span&gt;&lt;span class="o"&gt;);&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The coherence rule: &lt;strong&gt;one light direction for the whole page&lt;/strong&gt; (convention: above and slightly left), with vertical offset roughly 2× the horizontal. As elevation rises, offset and blur go up while opacity goes down. A card and a modal then read as the &lt;em&gt;same&lt;/em&gt; light hitting objects at different heights — not two unrelated effects. (Josh Comeau and Tobias Ahlin both have excellent deep-dives on this.)&lt;/p&gt;

&lt;p&gt;In dark mode, shadows nearly vanish, so switch to &lt;strong&gt;tonal elevation&lt;/strong&gt; — lighter surfaces for higher elevation — and never use pure &lt;code&gt;#000&lt;/code&gt; as your base (Material recommends &lt;code&gt;#121212&lt;/code&gt; with white overlays per level).&lt;/p&gt;




&lt;h2&gt;
  
  
  3. One accent, everything else greyscale
&lt;/h2&gt;

&lt;p&gt;Most "AI rainbow" UIs come from reaching for a new hue every time something needs emphasis. The discipline:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;One&lt;/strong&gt; accent for interactive emphasis.&lt;/li&gt;
&lt;li&gt;A full &lt;strong&gt;tinted-grey ramp&lt;/strong&gt; (not pure grey — nudge it 5–15% toward the brand hue) for text, surfaces, and borders.&lt;/li&gt;
&lt;li&gt;Only &lt;strong&gt;four&lt;/strong&gt; semantic colors — success / warning / error / info — used &lt;em&gt;strictly&lt;/em&gt; by meaning, never decoration.
&lt;/li&gt;
&lt;/ul&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight css"&gt;&lt;code&gt;&lt;span class="nd"&gt;:root&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="py"&gt;--accent&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="m"&gt;#5b5bd6&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="c"&gt;/* greys tinted slightly toward the accent's hue, not 0-saturation */&lt;/span&gt;
  &lt;span class="py"&gt;--grey-50&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;  &lt;span class="n"&gt;hsl&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="m"&gt;240&lt;/span&gt; &lt;span class="m"&gt;20%&lt;/span&gt; &lt;span class="m"&gt;98%&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
  &lt;span class="py"&gt;--grey-200&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;hsl&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="m"&gt;240&lt;/span&gt; &lt;span class="m"&gt;14%&lt;/span&gt; &lt;span class="m"&gt;90%&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
  &lt;span class="py"&gt;--grey-500&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;hsl&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="m"&gt;240&lt;/span&gt; &lt;span class="m"&gt;8%&lt;/span&gt;  &lt;span class="m"&gt;55%&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
  &lt;span class="py"&gt;--grey-900&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;hsl&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="m"&gt;240&lt;/span&gt; &lt;span class="m"&gt;12%&lt;/span&gt; &lt;span class="m"&gt;14%&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;  &lt;span class="c"&gt;/* body text — NOT #000 */&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;And one accessibility floor you don't get to negotiate (WCAG 2.2 AA): &lt;strong&gt;4.5:1&lt;/strong&gt; contrast for body text, &lt;strong&gt;3:1&lt;/strong&gt; for large text and UI components. Never convey information by color alone — pair it with an icon, text, or shape. A red border on an invalid field is not enough; add the icon and the message.&lt;/p&gt;




&lt;h2&gt;
  
  
  4. One spacing grid, with proximity doing the grouping
&lt;/h2&gt;

&lt;p&gt;Snap every margin, padding, and gap to a single scale — &lt;code&gt;4, 8, 12, 16, 24, 32, 48, 64&lt;/code&gt; — and let &lt;strong&gt;proximity&lt;/strong&gt; carry meaning. The rule from Gestalt (and Refactoring UI's "avoid ambiguous spacing"): the space &lt;em&gt;around&lt;/em&gt; a group must be clearly larger than the space &lt;em&gt;within&lt;/em&gt; it.&lt;/p&gt;

&lt;p&gt;A concrete ladder for forms:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;label → input        : 4–8px
input → input         : 12–16px
group → group (section): 24–32px
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;When everything is evenly spaced, the eye can't tell what belongs together. The doubling at each level is what makes a label bind to &lt;em&gt;its&lt;/em&gt; field and a section separate from the next.&lt;/p&gt;




&lt;h2&gt;
  
  
  Make the machine check itself
&lt;/h2&gt;

&lt;p&gt;Writing the decisions down is half the battle; the other half is catching violations. This is where AI can actually help — if you give it the rubric.&lt;/p&gt;

&lt;p&gt;I bake all of the above into an open-source design engine (&lt;a href="https://github.com/bitjaru/styleseed" rel="noopener noreferrer"&gt;StyleSeed&lt;/a&gt;) that Claude Code and Cursor read automatically, but the idea is tool-agnostic. The most useful piece turned out to be a &lt;strong&gt;coherence grader&lt;/strong&gt;: a check that scores a file and flags &lt;em&gt;mixed axes&lt;/em&gt; as real deductions.&lt;/p&gt;

&lt;p&gt;A trimmed example of what that output looks like:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight markdown"&gt;&lt;code&gt;&lt;span class="gu"&gt;## Design Score: 70 / 100   (Dashboard.tsx)&lt;/span&gt;

Color discipline   13/18  ▓▓▓░  #000 headings; 3 accent hues
Cards &amp;amp; elevation   8/12  ▓▓░░  1px borders doing separation
Coherence           6/12  ▓▓░░  sharp cards (l.22) + pill buttons (l.48); 3 accents

&lt;span class="gu"&gt;### Fix first (highest score gain)&lt;/span&gt;
&lt;span class="p"&gt;1.&lt;/span&gt; Unify radius (pick soft 8–12px) + collapse to one accent  → +9
&lt;span class="p"&gt;2.&lt;/span&gt; Add empty + loading states to the orders list            → +7
&lt;span class="p"&gt;3.&lt;/span&gt; Drop 1px borders, use tone + ≤8% shadow                  → +4
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The "Coherence" category is the one that most predicts "looks AI-generated," precisely because it measures &lt;em&gt;system-wide consistency&lt;/em&gt; rather than per-component prettiness. A component can be beautiful and still wrong — if it disagrees with its neighbors.&lt;/p&gt;




&lt;h2&gt;
  
  
  The takeaway
&lt;/h2&gt;

&lt;p&gt;If you remember one thing:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Beautiful parts don't make a beautiful UI. &lt;strong&gt;Agreeing&lt;/strong&gt; parts do.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Pick one value per axis — radius, shadow, accent, spacing, icons, type, motion — write it down as a token, and make every new element inherit it instead of inventing a fresh choice. That single discipline is the difference between "a robot made this" and "a designer made this," and it's the cheapest, highest-leverage thing you can do for an AI-built interface.&lt;/p&gt;

&lt;p&gt;The corners are sharp? Then everything is sharp. All the way down.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Want the practical side? I covered &lt;a href="https://dev.to/kiwibreaksme/your-vibe-coded-app-looks-ugly-heres-the-one-file-fix-4ef0"&gt;the one-file fix that makes your vibe-coded app stop looking ugly&lt;/a&gt; in a companion post. And the full research-backed version — concrete spacing numbers, a type recipe per app type, layered-shadow and nested-radius recipes, all grounded in Refactoring UI / Material 3 / Apple HIG / WCAG — is open source (MIT): &lt;a href="https://github.com/bitjaru/styleseed" rel="noopener noreferrer"&gt;github.com/bitjaru/styleseed&lt;/a&gt;. A ⭐ helps more devs (and more AI tools) find it.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>design</category>
      <category>css</category>
      <category>webdev</category>
      <category>ai</category>
    </item>
    <item>
      <title>Your vibe-coded app looks ugly. Here's the one-file fix.</title>
      <dc:creator>@kiwibreaksme</dc:creator>
      <pubDate>Tue, 14 Apr 2026 03:59:47 +0000</pubDate>
      <link>https://dev.to/kiwibreaksme/your-vibe-coded-app-looks-ugly-heres-the-one-file-fix-4ef0</link>
      <guid>https://dev.to/kiwibreaksme/your-vibe-coded-app-looks-ugly-heres-the-one-file-fix-4ef0</guid>
      <description>&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F2yx5kvlddc5pmpysgz21.gif" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F2yx5kvlddc5pmpysgz21.gif" alt="StyleSeed live demo" width="600" height="817"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Let's be honest about vibe coding
&lt;/h2&gt;

&lt;p&gt;You tell Claude Code "build me a SaaS dashboard." Thirty seconds later you have a working app. Revenue chart, user table, activity feed. Everything functional.&lt;/p&gt;

&lt;p&gt;And it looks like &lt;strong&gt;every other AI-generated app on the internet&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;Spacing that doesn't breathe. Seven accent colors fighting for attention. Cards that float on the page with zero separation from the background. The components are all shadcn — the technical stack is fine. But the &lt;em&gt;result&lt;/em&gt; is amateur.&lt;/p&gt;

&lt;p&gt;I spent a week trying to fix this through prompts. "Use only one accent color." "Card background &lt;code&gt;#FFF&lt;/code&gt;, page background &lt;code&gt;#FAFAFA&lt;/code&gt;." "Smaller shadows." Every fix worked for one screen, then the next screen forgot.&lt;/p&gt;

&lt;p&gt;Then it clicked: &lt;strong&gt;the model doesn't need better prompts. It needs a design brain.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Pro designers aren't using better components — they're using invisible rules
&lt;/h2&gt;

&lt;p&gt;Ask a senior designer why their dashboard looks refined and they'll say things like:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"Never use pure black. Use &lt;code&gt;#2A2A2A&lt;/code&gt;."&lt;/p&gt;

&lt;p&gt;"One accent color per app. Everything else grayscale."&lt;/p&gt;

&lt;p&gt;"Shadows at 4% opacity. If you can see it, it's already too much."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Nobody writes these down. They're baked into years of experience, invisible to outsiders — which means invisible to LLMs. No matter how many shadcn components Claude has in its training data, it has never been told &lt;em&gt;when&lt;/em&gt; to use which.&lt;/p&gt;

&lt;p&gt;So I sat down and extracted 69 of these rules from Toss, Stripe, Linear, and Vercel — the brands whose output I wanted Claude to imitate — and wrote them into a single &lt;code&gt;DESIGN-LANGUAGE.md&lt;/code&gt; file that Claude reads automatically.&lt;/p&gt;

&lt;h2&gt;
  
  
  The 69 rules, organized
&lt;/h2&gt;

&lt;p&gt;The rules cluster into six groups:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Color discipline&lt;/strong&gt; (12 rules) — &lt;code&gt;#2A2A2A&lt;/code&gt; as the refined black, 5-level grayscale (&lt;code&gt;#2A → #3C → #6A → #7A → #9B&lt;/code&gt;), one accent color maximum, no pure white on pure white.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Spatial rhythm&lt;/strong&gt; (14 rules) — never repeat the same section type consecutively, alternate tall/compact, &lt;code&gt;2:1&lt;/code&gt; number-to-unit ratios (&lt;code&gt;48px&lt;/code&gt; value, &lt;code&gt;24px&lt;/code&gt; unit).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Information hierarchy&lt;/strong&gt; (9 rules) — card/background separation (&lt;code&gt;#FFF&lt;/code&gt; card on &lt;code&gt;#FAFAFA&lt;/code&gt; page) matters more than any border, density &lt;em&gt;increases&lt;/em&gt; as you scroll down, top is big numbers, bottom is dense lists.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Shadow &amp;amp; elevation&lt;/strong&gt; (8 rules) — max &lt;code&gt;4%&lt;/code&gt; opacity, elevation is never &lt;code&gt;z-index&lt;/code&gt;, dark mode replaces shadows with borders.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Component variance&lt;/strong&gt; (11 rules) — never 4 identical KPI cards, stagger content types (2 with trend, 1 with progress, 1 with comparison), avoid "filled form of sameness."&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Motion &amp;amp; feedback&lt;/strong&gt; (15 rules) — &lt;code&gt;200ms&lt;/code&gt; normal, spring for entrance, &lt;code&gt;ease-out&lt;/code&gt; for exit, hover lifts by &lt;code&gt;6px&lt;/code&gt; max, tap scales to &lt;code&gt;0.98&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Drop the file into a project, Claude reads it, Cursor reads it (via &lt;code&gt;.cursorrules&lt;/code&gt;). Same prompt, same model — the output quality jumps.&lt;/p&gt;

&lt;h2&gt;
  
  
  One rule, unpacked
&lt;/h2&gt;

&lt;p&gt;Take the &lt;code&gt;#2A2A2A&lt;/code&gt; rule. Pure &lt;code&gt;#000&lt;/code&gt; on &lt;code&gt;#FFF&lt;/code&gt; has a contrast ratio of 21:1 — the highest possible. Sounds like a win. But at that contrast the eye gets fatigued fast, and the text starts to feel harsh and heavy on a white page.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;#2A2A2A&lt;/code&gt; drops the contrast to about 15:1. Still WCAG AAA for body text. But the page stops "vibrating." Every refined design system in the last decade — Apple, Linear, Vercel, Notion, Toss — uses a softened near-black for body text. Claude picks &lt;code&gt;#000&lt;/code&gt; by default because it's the highest-contrast option, which is technically correct and aesthetically wrong.&lt;/p&gt;

&lt;p&gt;One rule. Fixes thousands of screens.&lt;/p&gt;

&lt;p&gt;Now multiply that by 69.&lt;/p&gt;

&lt;h2&gt;
  
  
  Same component, three brand DNAs
&lt;/h2&gt;

&lt;p&gt;The live demo up top is a real app — &lt;a href="https://styleseed-demo.vercel.app" rel="noopener noreferrer"&gt;styleseed-demo.vercel.app&lt;/a&gt;. Click the Toss / Raycast / Arc switcher. Same chat component, three totally different brand identities — colors, radius, shadows, motion durations, gradients, even the phone frame's ambient light.&lt;/p&gt;

&lt;p&gt;The switcher flips a single &lt;code&gt;data-skin&lt;/code&gt; attribute on the wrapper. That's it. No conditional rendering, no theme provider, no re-mount.&lt;/p&gt;

&lt;h2&gt;
  
  
  How it actually works
&lt;/h2&gt;

&lt;p&gt;Every visual value is a CSS variable scoped to &lt;code&gt;[data-skin="..."]&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight css"&gt;&lt;code&gt;&lt;span class="o"&gt;[&lt;/span&gt;&lt;span class="nt"&gt;data-skin&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s1"&gt;"toss"&lt;/span&gt;&lt;span class="o"&gt;]&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="py"&gt;--brand&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="m"&gt;#3182F6&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="py"&gt;--radius&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="m"&gt;0.875rem&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="py"&gt;--shadow-card&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="m"&gt;0&lt;/span&gt; &lt;span class="m"&gt;1px&lt;/span&gt; &lt;span class="m"&gt;3px&lt;/span&gt; &lt;span class="n"&gt;rgba&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="m"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="m"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="m"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="m"&gt;0.04&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
  &lt;span class="py"&gt;--gradient-brand&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;linear-gradient&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="m"&gt;135deg&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="m"&gt;#3182F6&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="m"&gt;#4A90F7&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
  &lt;span class="py"&gt;--duration-normal&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="m"&gt;200ms&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;

&lt;span class="o"&gt;[&lt;/span&gt;&lt;span class="nt"&gt;data-skin&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s1"&gt;"raycast"&lt;/span&gt;&lt;span class="o"&gt;]&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="py"&gt;--brand&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="m"&gt;#FF4E8B&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="py"&gt;--radius&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="m"&gt;0.5rem&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="py"&gt;--shadow-card&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="m"&gt;0&lt;/span&gt; &lt;span class="m"&gt;0&lt;/span&gt; &lt;span class="m"&gt;0&lt;/span&gt; &lt;span class="m"&gt;1px&lt;/span&gt; &lt;span class="n"&gt;rgba&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="m"&gt;255&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="m"&gt;255&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="m"&gt;255&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="m"&gt;0.05&lt;/span&gt;&lt;span class="p"&gt;),&lt;/span&gt;
                 &lt;span class="m"&gt;0&lt;/span&gt; &lt;span class="m"&gt;8px&lt;/span&gt; &lt;span class="m"&gt;24px&lt;/span&gt; &lt;span class="n"&gt;rgba&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="m"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="m"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="m"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="m"&gt;0.4&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
  &lt;span class="py"&gt;--gradient-brand&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;linear-gradient&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="m"&gt;135deg&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="m"&gt;#FF6363&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="m"&gt;#FF8E3C&lt;/span&gt; &lt;span class="m"&gt;30%&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
                    &lt;span class="m"&gt;#E84A8E&lt;/span&gt; &lt;span class="m"&gt;65%&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="m"&gt;#A855F7&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
  &lt;span class="py"&gt;--duration-normal&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="m"&gt;220ms&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Components never touch the raw values — only the tokens:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight tsx"&gt;&lt;code&gt;&lt;span class="p"&gt;&amp;lt;&lt;/span&gt;&lt;span class="nt"&gt;motion&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nt"&gt;div&lt;/span&gt;
  &lt;span class="na"&gt;style&lt;/span&gt;&lt;span class="p"&gt;=&lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="na"&gt;background&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;var(--card)&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="na"&gt;borderRadius&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;var(--radius-xl)&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="na"&gt;boxShadow&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;var(--shadow-card)&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;/&amp;gt;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Swap the data attribute → the entire tree morphs in one frame. Framer Motion animates the transitions for free because the underlying CSS properties are inherited.&lt;/p&gt;

&lt;p&gt;This is the trick the 69 rules depend on. Rules reference &lt;em&gt;semantic&lt;/em&gt; tokens (&lt;code&gt;--brand&lt;/code&gt;, &lt;code&gt;--card&lt;/code&gt;, &lt;code&gt;--shadow-card&lt;/code&gt;), not literal values. So the same rulebook works whether your app looks like Toss or Vercel or your client's weird purple brand.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why this matters for vibe coders
&lt;/h2&gt;

&lt;p&gt;The point of vibe coding is shipping fast. The problem with vibe coding is the output screams "I was made in 30 minutes."&lt;/p&gt;

&lt;p&gt;StyleSeed closes that gap in two moves:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;The 69 rules fix the judgment.&lt;/strong&gt; Claude stops picking &lt;code&gt;#000&lt;/code&gt; for text and &lt;code&gt;py-4&lt;/code&gt; for everything. The output starts looking designed, not generated.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The token structure fixes the brand.&lt;/strong&gt; Your MVP can ship with Toss skin today and re-skin to a client's brand tomorrow by swapping one file. Same codebase. Zero rewrites.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Both drop in with one command. Engine is brand-agnostic — the rules don't know what color your brand is.&lt;/p&gt;

&lt;h2&gt;
  
  
  Stack
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Next.js 16&lt;/strong&gt; (App Router, Turbopack)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Tailwind CSS v4&lt;/strong&gt; (the &lt;code&gt;@theme inline&lt;/code&gt; syntax makes token sharing trivial)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Framer Motion&lt;/strong&gt; for the morph animation&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Radix&lt;/strong&gt; under the components&lt;/li&gt;
&lt;li&gt;Free, MIT&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;🎬 &lt;strong&gt;Live demo:&lt;/strong&gt; &lt;a href="https://styleseed-demo.vercel.app" rel="noopener noreferrer"&gt;styleseed-demo.vercel.app&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;⭐ &lt;strong&gt;GitHub:&lt;/strong&gt; &lt;a href="https://github.com/bitjaru/styleseed" rel="noopener noreferrer"&gt;bitjaru/styleseed&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;📚 11 Claude Code slash commands included (&lt;code&gt;/ss-page&lt;/code&gt;, &lt;code&gt;/ss-review&lt;/code&gt;, &lt;code&gt;/ss-audit&lt;/code&gt;, ...). Works with Cursor.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you're vibe coding anything right now, drop the engine in and compare the output. Honest critique on the rules themselves very welcome — which one feels wrong? Which one's missing? The components are downstream. &lt;strong&gt;The rules are the actual product.&lt;/strong&gt;&lt;/p&gt;

</description>
      <category>webdev</category>
      <category>ai</category>
      <category>vibecoding</category>
      <category>showdev</category>
    </item>
    <item>
      <title>Your Vibe-Coded App Looks Ugly. Here's What I Did About It.</title>
      <dc:creator>@kiwibreaksme</dc:creator>
      <pubDate>Thu, 09 Apr 2026 06:26:23 +0000</pubDate>
      <link>https://dev.to/kiwibreaksme/your-vibe-coded-app-looks-ugly-heres-what-i-did-about-it-2nb4</link>
      <guid>https://dev.to/kiwibreaksme/your-vibe-coded-app-looks-ugly-heres-what-i-did-about-it-2nb4</guid>
      <description>&lt;p&gt;Again. Not pretty.&lt;/p&gt;

&lt;p&gt;I asked Claude Code to build a dashboard. It worked perfectly. And it looked like a hackathon project.&lt;/p&gt;

&lt;p&gt;Card spacing was random. Typography had no hierarchy. Colors were everywhere. The whole thing felt... off.&lt;/p&gt;

&lt;p&gt;I tried "make it more professional." Didn't help. Tried "like the Toss app." Slightly better, but still amateur.&lt;/p&gt;

&lt;p&gt;The problem wasn't the AI. The AI makes great components. It just can't make &lt;strong&gt;design decisions&lt;/strong&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Mean
&lt;/h2&gt;

&lt;p&gt;Open the Toss app (Korea's biggest fintech) and look closely. Every screen feels the same. Not boring-same — &lt;em&gt;cohesive&lt;/em&gt;-same. I spent a while figuring out why.&lt;/p&gt;

&lt;p&gt;There are rules.&lt;/p&gt;

&lt;p&gt;The darkest black isn't &lt;code&gt;#000&lt;/code&gt;. It's &lt;code&gt;#2A2A2A&lt;/code&gt;. Tiny difference. But that's the gap between "harsh" and "refined."&lt;/p&gt;

&lt;p&gt;There's exactly one accent color in the entire app. One. Everything else is grayscale. Restraint is what makes it feel expensive.&lt;/p&gt;

&lt;p&gt;Card shadows? 4% opacity. If you can see the shadow, it's already too much.&lt;/p&gt;

&lt;p&gt;Numbers at 48px, units at 24px. Always 2:1. Your eye hits the number first.&lt;/p&gt;

&lt;p&gt;Designers make these calls instinctively. Years of experience. But here's the thing — if you write them down, AI can follow them too.&lt;/p&gt;

&lt;h2&gt;
  
  
  So I Wrote Them Down
&lt;/h2&gt;

&lt;p&gt;I was building my own app and got tired of fixing ugly AI output. Started asking myself "why does this look bad?" for every screen. Patterns emerged. Same mistakes, every time.&lt;/p&gt;

&lt;p&gt;So I documented the judgment calls a designer would make. Ended up with 69 rules.&lt;/p&gt;

&lt;p&gt;Like:&lt;/p&gt;

&lt;p&gt;Never use the same card layout twice in a row. Hero, then grid, then chart, then list. There needs to be rhythm.&lt;/p&gt;

&lt;p&gt;When you make 4 KPI cards, don't make them identical. Two with trend arrows, one with a progress bar, one with comparison text. Variation keeps it from feeling robotic.&lt;/p&gt;

&lt;p&gt;As you scroll down, text gets smaller and information gets denser. Hero is one big 48px number. Bottom of the page is 14px detail rows. Natural zoom-in effect.&lt;/p&gt;

&lt;p&gt;I packaged these rules so AI could read them.&lt;/p&gt;

&lt;h2&gt;
  
  
  Engine + Skins
&lt;/h2&gt;

&lt;p&gt;The structure is kind of fun.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;styleseed/
├── engine/    ← Design brain (69 rules, 48 components, 11 AI skills)
└── skins/     ← Brand colors (Toss, Stripe, Linear, Vercel, Notion...)
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The &lt;strong&gt;engine&lt;/strong&gt; knows &lt;em&gt;how to structure&lt;/em&gt; your app. Layout, typography ratios, card structure, visual rhythm, forbidden patterns. Doesn't care what brand you are.&lt;/p&gt;

&lt;p&gt;A &lt;strong&gt;skin&lt;/strong&gt; is one CSS file. &lt;code&gt;--brand: #533afd&lt;/code&gt; makes it Stripe. &lt;code&gt;--brand: #721FE5&lt;/code&gt; makes it Toss. Change one variable, whole UI follows.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://github.com/VoltAgent/awesome-design-md" rel="noopener noreferrer"&gt;awesome-design-md&lt;/a&gt; has 58 brand identities you can grab as skins.&lt;/p&gt;

&lt;h2&gt;
  
  
  How I Actually Use It
&lt;/h2&gt;

&lt;p&gt;Step 1. Copy the engine.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;cp&lt;/span&gt; &lt;span class="nt"&gt;-r&lt;/span&gt; styleseed/engine/&lt;span class="k"&gt;*&lt;/span&gt; my-project/
&lt;span class="nb"&gt;cp&lt;/span&gt; &lt;span class="nt"&gt;-r&lt;/span&gt; styleseed/engine/css/&lt;span class="k"&gt;*&lt;/span&gt; my-project/src/styles/
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Step 2. Pick a skin.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;cp &lt;/span&gt;styleseed/skins/stripe/theme.css my-project/src/styles/theme.css
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Or run &lt;code&gt;/ui-setup&lt;/code&gt; and it walks you through everything — app type, colors, font, generates your first page.&lt;/p&gt;

&lt;p&gt;Step 3. Build.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Build me a SaaS dashboard.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;AI reads &lt;code&gt;CLAUDE.md&lt;/code&gt; automatically. Follows the 69 rules. Lays out hero card, KPI grid, chart section, activity list in information-pyramid order.&lt;/p&gt;

&lt;p&gt;But it doesn't stop there.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;/ux-audit src/Dashboard.tsx    &lt;span class="c"&gt;# Nielsen's heuristics check&lt;/span&gt;
/ux-copy &lt;span class="s2"&gt;"dashboard"&lt;/span&gt;           &lt;span class="c"&gt;# Auto-generate button labels, error messages&lt;/span&gt;
/ux-feedback src/Dashboard.tsx  &lt;span class="c"&gt;# Add loading/error/empty states&lt;/span&gt;
/ui-review src/Dashboard.tsx    &lt;span class="c"&gt;# Design compliance check&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Four commands that do what a designer + UX researcher would do. Felt like cheating, honestly.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why It Feels Cohesive
&lt;/h2&gt;

&lt;p&gt;Why does the Toss app feel like "one app" no matter what screen you're on?&lt;/p&gt;

&lt;p&gt;Because one team's consistent judgment is applied everywhere. "This spacing is 24px." "Don't use that color." "Numbers are this size." Same rules, every screen, every time.&lt;/p&gt;

&lt;p&gt;StyleSeed is those rules, written as 69 lines of design judgment. AI reads them. Every page it builds follows the same standard.&lt;/p&gt;

&lt;p&gt;Swap the skin — Stripe, Vercel, whatever — the structure stays consistent. Same engine.&lt;/p&gt;

&lt;h2&gt;
  
  
  Honestly
&lt;/h2&gt;

&lt;p&gt;It's not perfect. Mobile-only (430px) for now. Desktop layouts need separate work. Charts need manual Recharts setup. I'm still adding rules and refining.&lt;/p&gt;

&lt;p&gt;But the difference between before and after is obvious. "It looks bad but I don't know how to fix it" turned into "it just comes out clean."&lt;/p&gt;

&lt;p&gt;No designer. Five minutes of setup.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://github.com/bitjaru/styleseed" class="crayons-btn crayons-btn--primary" rel="noopener noreferrer"&gt;Star on GitHub&lt;/a&gt;
&lt;/p&gt;

&lt;p&gt;MIT open source. Built for my own app, sharing because it might save you the same headache.&lt;/p&gt;

</description>
      <category>vibecoding</category>
      <category>claude</category>
      <category>ai</category>
      <category>opensource</category>
    </item>
    <item>
      <title>디자이너 없이 바이브코딩으로 토스급 UI 만드는 법</title>
      <dc:creator>@kiwibreaksme</dc:creator>
      <pubDate>Thu, 09 Apr 2026 06:25:56 +0000</pubDate>
      <link>https://dev.to/kiwibreaksme/dijaineo-eobsi-baibeukodingeuro-toseugeub-ui-mandeuneun-beob-5c0g</link>
      <guid>https://dev.to/kiwibreaksme/dijaineo-eobsi-baibeukodingeuro-toseugeub-ui-mandeuneun-beob-5c0g</guid>
      <description>&lt;p&gt;또. 이쁘지가 않다.&lt;/p&gt;

&lt;p&gt;Claude Code한테 "대시보드 만들어줘" 했더니 기능은 되는데 화면이 영 아마추어다. 카드 간격이 제각각이고, 색이 정리가 안 되고, 전체적으로 뭔가... 촌스럽다.&lt;/p&gt;

&lt;p&gt;"좀 더 깔끔하게" 해봤자 달라지는 게 없다. 좀 더 구체적으로 "토스처럼" 해봤자 비슷하다. AI가 토스를 모르는 게 아니다. 컴포넌트도 잘 만든다.&lt;/p&gt;

&lt;p&gt;근데 &lt;strong&gt;디자인 판단&lt;/strong&gt;을 못 한다.&lt;/p&gt;

&lt;h2&gt;
  
  
  무슨 말이냐면
&lt;/h2&gt;

&lt;p&gt;토스 앱을 켜서 찬찬히 보면, 화면마다 같은 "느낌"이 있다. 그 느낌의 정체가 뭔지 뜯어봤다.&lt;/p&gt;

&lt;p&gt;규칙이 있다.&lt;/p&gt;

&lt;p&gt;가장 어두운 검정이 &lt;code&gt;#000&lt;/code&gt;이 아니라 &lt;code&gt;#2A2A2A&lt;/code&gt;다. 미세한 차인데, 이게 "딱딱함"과 "고급스러움"의 차이다.&lt;/p&gt;

&lt;p&gt;강조 색은 앱 전체에서 딱 하나. 보라 하나. 나머지는 전부 회색 톤이다. 절제가 고급이다.&lt;/p&gt;

&lt;p&gt;카드 그림자? opacity 4%다. 그림자가 눈에 보이면 이미 과하다.&lt;/p&gt;

&lt;p&gt;숫자는 48px로 크게, 단위는 24px로 작게. 항상 2:1. 시선이 숫자에 먼저 간다.&lt;/p&gt;

&lt;p&gt;디자이너는 이런 판단을 무의식적으로 내린다. 경험으로 쌓인 거다. 근데 이걸 글로 정리하면? AI도 따라할 수 있다.&lt;/p&gt;

&lt;h2&gt;
  
  
  그래서 정리했다
&lt;/h2&gt;

&lt;p&gt;내 앱 만들면서 AI가 뱉는 UI를 고치다 지쳤다. "이건 왜 구려 보이지?"를 분석하다 보니 패턴이 보였다. 같은 실수가 반복됐다.&lt;/p&gt;

&lt;p&gt;그래서 디자이너가 머릿속에 갖고 있는 판단 기준을 하나씩 글로 적었다. 69개가 됐다.&lt;/p&gt;

&lt;p&gt;예를 들면:&lt;/p&gt;

&lt;p&gt;같은 카드 레이아웃을 연속으로 두 번 쓰지 마라. 히어로 다음에 그리드, 그리드 다음에 차트, 차트 다음에 리스트. 리듬이 있어야 한다.&lt;/p&gt;

&lt;p&gt;KPI 카드 4개를 만들 때 전부 똑같이 만들지 마라. 트렌드 화살표 2개, 프로그레스바 1개, 비교 텍스트 1개. 변주가 있어야 단조롭지 않다.&lt;/p&gt;

&lt;p&gt;위에서 아래로 갈수록 글씨가 작아지고 정보가 빽빽해진다. 히어로는 48px 하나, 맨 아래 리스트는 14px 여러 줄. 자연스러운 줌인 효과.&lt;/p&gt;

&lt;p&gt;이 규칙들을 AI가 읽을 수 있게 패키징했다.&lt;/p&gt;

&lt;h2&gt;
  
  
  엔진 + 스킨
&lt;/h2&gt;

&lt;p&gt;구조가 좀 재밌다.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;styleseed/
├── engine/    ← 디자인 두뇌 (규칙 69개, 컴포넌트 48개, AI 스킬 11개)
└── skins/     ← 브랜드 컬러 (토스, Stripe, Linear, Vercel, Notion...)
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;엔진&lt;/strong&gt;은 "어떻게 구성할지"를 안다. 레이아웃, 타이포 비율, 카드 구조, 시각적 리듬, 금지 패턴. 브랜드 색이 뭐든 상관없이 동작한다.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;스킨&lt;/strong&gt;은 CSS 파일 하나다. &lt;code&gt;--brand: #533afd&lt;/code&gt; 하면 Stripe, &lt;code&gt;--brand: #721FE5&lt;/code&gt; 하면 토스. 색만 바꾸면 전체 UI가 따라 바뀐다.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://github.com/VoltAgent/awesome-design-md" rel="noopener noreferrer"&gt;awesome-design-md&lt;/a&gt;에 58개 브랜드가 있는데, 거기서 아무거나 골라서 스킨으로 쓸 수 있다.&lt;/p&gt;

&lt;h2&gt;
  
  
  실제로 쓰는 과정
&lt;/h2&gt;

&lt;p&gt;1단계. 엔진을 내 프로젝트에 복사한다.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;cp&lt;/span&gt; &lt;span class="nt"&gt;-r&lt;/span&gt; styleseed/engine/&lt;span class="k"&gt;*&lt;/span&gt; my-project/
&lt;span class="nb"&gt;cp&lt;/span&gt; &lt;span class="nt"&gt;-r&lt;/span&gt; styleseed/engine/css/&lt;span class="k"&gt;*&lt;/span&gt; my-project/src/styles/
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;2단계. 스킨을 고른다.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;cp &lt;/span&gt;styleseed/skins/stripe/theme.css my-project/src/styles/theme.css
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;귀찮으면 &lt;code&gt;/ui-setup&lt;/code&gt; 치면 하나씩 물어보면서 세팅해준다. 앱 종류, 색상, 폰트, 첫 페이지까지.&lt;/p&gt;

&lt;p&gt;3단계. 만든다.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;SaaS 대시보드 만들어줘.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;AI가 &lt;code&gt;CLAUDE.md&lt;/code&gt;를 자동으로 읽는다. 69개 규칙에 따라 히어로 카드, KPI 그리드, 차트 섹션, 리스트를 정보 피라미드 순서로 배치한다.&lt;/p&gt;

&lt;p&gt;여기서 끝이 아니다.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;/ux-audit src/Dashboard.tsx    &lt;span class="c"&gt;# 닐슨 10대 원칙으로 UX 점검&lt;/span&gt;
/ux-copy &lt;span class="s2"&gt;"대시보드"&lt;/span&gt;             &lt;span class="c"&gt;# 버튼 라벨, 에러 메시지 자동 생성&lt;/span&gt;
/ux-feedback src/Dashboard.tsx  &lt;span class="c"&gt;# 로딩/에러/빈 상태 추가&lt;/span&gt;
/ui-review src/Dashboard.tsx    &lt;span class="c"&gt;# 디자인 규칙 준수 검사&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;솔직히 좀 사기다 싶었다. 디자이너 + UX 리서처가 하는 일을 명령어 4개로.&lt;/p&gt;

&lt;h2&gt;
  
  
  통일감의 비결
&lt;/h2&gt;

&lt;p&gt;왜 토스 앱은 어떤 화면을 열어도 "같은 앱" 느낌이 날까.&lt;/p&gt;

&lt;p&gt;한 사람(또는 한 팀)의 머릿속에 일관된 기준이 있기 때문이다. "이 간격은 24px", "이 색은 쓰지 마", "숫자는 이 크기" — 이 기준이 모든 화면에 적용되니까 통일감이 생긴다.&lt;/p&gt;

&lt;p&gt;StyleSeed는 그 기준을 69개 규칙으로 적어놓은 거다. AI가 이걸 읽으면, 어떤 페이지를 만들든 같은 기준으로 만든다.&lt;/p&gt;

&lt;p&gt;스킨이 바뀌어도 구조는 안 바뀐다. Stripe든 Vercel이든 — 엔진이 동일하니까.&lt;/p&gt;

&lt;h2&gt;
  
  
  솔직히 말하면
&lt;/h2&gt;

&lt;p&gt;완벽하지는 않다. 아직 모바일(430px) 전용이고, 데스크톱 레이아웃은 별도로 짜야 한다. 차트도 Recharts 직접 세팅해야 한다. 계속 규칙 추가하면서 다듬는 중이다.&lt;/p&gt;

&lt;p&gt;근데 확실한 건 — 이거 쓰기 전과 후가 확연히 다르다. "구려 보이는데 뭘 고쳐야 할지 모르겠다"에서 "알아서 깔끔하게 나온다"로.&lt;/p&gt;

&lt;p&gt;디자이너 고용비 0원. 세팅 5분.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://github.com/bitjaru/styleseed" class="crayons-btn crayons-btn--primary" rel="noopener noreferrer"&gt;GitHub에서 보기&lt;/a&gt;
&lt;/p&gt;

&lt;p&gt;MIT 오픈소스다. 쓸만하면 스타 하나 부탁.&lt;/p&gt;

</description>
      <category>vibecoding</category>
      <category>claude</category>
      <category>ai</category>
      <category>오픈소스</category>
    </item>
    <item>
      <title>당신이 AI를 잘 못 쓰는 이유, 메타인지 훈련</title>
      <dc:creator>@kiwibreaksme</dc:creator>
      <pubDate>Thu, 26 Feb 2026 04:55:20 +0000</pubDate>
      <link>https://dev.to/kiwibreaksme/dangsini-aireul-jal-mos-sseuneun-iyu-05co-4b0c</link>
      <guid>https://dev.to/kiwibreaksme/dangsini-aireul-jal-mos-sseuneun-iyu-05co-4b0c</guid>
      <description>&lt;p&gt;&lt;em&gt;AI랑 일 잘하는 법 — 1편&lt;/em&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;먼저 말해두면, 이 글은 프롬프트 템플릿 모음이 아니다. "이렇게 물어보세요~" 하는 글은 이미 넘쳐난다. 이 글은 그보다 한 단계 아래에 있는 얘기를 한다. &lt;strong&gt;생각하는 방식&lt;/strong&gt; 얘기다. 그리고 이게 바뀌면 AI 결과물만 좋아지는 게 아니라, 본인이 성장한다. 진짜로.&lt;/p&gt;
&lt;/blockquote&gt;




&lt;p&gt;지난주에 팀 회식을 해야 했다.&lt;/p&gt;

&lt;p&gt;귀찮아서 AI한테 물어봤다. "강남역 근처 고깃집 추천해줘."&lt;/p&gt;

&lt;p&gt;5곳을 뽑아줬다. 리뷰 좋고, 분위기 좋고, 가격도 적당하다고 했다. 예약 팁까지 줬다. 친절하다.&lt;/p&gt;

&lt;p&gt;첫 번째 가게를 네이버에 검색해봤다. &lt;strong&gt;3개월 전에 폐업했다.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;두 번째는 이름이 바뀌어 있었다. 세 번째는 별 1개 테러를 맞고 있었다.&lt;/p&gt;

&lt;p&gt;동료한테 얘기했더니 반응이 두 가지였다.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"역시 AI는 쓸모없네."&lt;br&gt;
"그걸 왜 AI한테 물어봐 ㅋㅋ"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;둘 다 핵심을 모르는 거다.&lt;/p&gt;

&lt;p&gt;AI가 문제가 아니다. &lt;strong&gt;물어보는 방식이 문제다.&lt;/strong&gt;&lt;/p&gt;




&lt;h3&gt;
  
  
  그러면 어떻게 물어봤어야 하나
&lt;/h3&gt;

&lt;p&gt;애초에 "고깃집 추천해줘"가 잘못된 질문이다.&lt;/p&gt;

&lt;h4&gt;
  
  
  ❌ 보통 사람의 질문
&lt;/h4&gt;

&lt;blockquote&gt;
&lt;p&gt;"강남역 근처 회식하기 좋은 고깃집 추천해줘"&lt;br&gt;
→ AI가 5곳 추천 → 첫 번째가 폐업 → "AI 별로네"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h4&gt;
  
  
  ✅ 이렇게 물어봤어야 한다
&lt;/h4&gt;

&lt;p&gt;&lt;strong&gt;1단계 — 기준부터 잡기:&lt;/strong&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"회사 회식할 때 직원들 만족도가 높은 음식점 유형이 뭐야? 인원수, 분위기, 메뉴 선택지 같은 요소별로 정리해줘."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;AI가 판단 기준을 먼저 정리해준다. 고깃집이 무난하긴 한데, 채식하는 사람 있으면 한정식이 낫고, 10명 넘으면 룸 있는 데가 좋고. 이런 거.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2단계 — 조건 넣기:&lt;/strong&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"우리 팀 8명이고, 고기파 많고, 강남역 도보 10분 이내. 후보 뽑아줘"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;&lt;strong&gt;3단계 — 결과는 내가 검증:&lt;/strong&gt;&lt;br&gt;
네이버 지도에서 영업 중인지, 리뷰는 어떤지 직접 확인한다. 이건 사람이 해야 하는 영역이다.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;(더 미친 사람은 AI한테 네이버 지도 데이터를 직접 물려줘서, 영업 중인 가게만 자동으로 걸러내게 한다. 이것도 나중에 다룬다.)&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;"추천해줘" 한 마디로 끝내는 거랑, 기준 잡고 → 조건 넣고 → 결과를 내가 검증하는 거랑은 &lt;strong&gt;완전히 다른 행위다.&lt;/strong&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;💡 여기서 한 가지 더. 반대로 생각하면 재밌다. 사람들이 구글 대신 AI한테 "추천해줘"를 점점 더 많이 하고 있다. 가게 입장에서는? 구글에 잘 걸리는 게 중요했는데, 이제는 &lt;strong&gt;AI가 추천하는 리스트에 들어가는 게 중요해졌다.&lt;/strong&gt; 이걸 'AI SEO'라고 부르기 시작했다. 이것도 다른 편에서 다룬다.&lt;/p&gt;
&lt;/blockquote&gt;


&lt;h2&gt;
  
  
  대부분이 이렇게 쓰고 있다
&lt;/h2&gt;

&lt;p&gt;ChatGPT, Claude, Gemini. 한 번쯤은 써봤을 거다.&lt;/p&gt;

&lt;p&gt;근데 대부분 이렇게 쓴다.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;"OO해줘." → 결과 받음 → 끝.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;자기소개서 써줘. 이메일 번역해줘. 여행 계획 세워줘. 한 번 물어보고 한 번 받고 끝이다.&lt;/p&gt;

&lt;p&gt;좋으면 "AI 대단하다." 별로면 "AI 별로네." 이게 전부다.&lt;/p&gt;

&lt;p&gt;이건 구글 검색이랑 같은 거다. 검색창에 키워드 넣고 첫 번째 결과 클릭하는 거.&lt;/p&gt;

&lt;p&gt;AI는 검색엔진이 아니다. 전지전능한 신은 더더욱 아니다.&lt;/p&gt;

&lt;p&gt;근데 신한테 소원 빌듯이 한 줄 던지고 기적을 기대한다.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;기적은 안 일어난다.&lt;/strong&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  Meta의 AI 안전 전문가도 당했다
&lt;/h2&gt;

&lt;p&gt;최근에 딱 맞는 사건이 있었다.&lt;/p&gt;

&lt;p&gt;요즘 IT업계에서 제일 핫한 게 &lt;strong&gt;OpenClaw&lt;/strong&gt;라는 AI 도구다. 내 컴퓨터에서 직접 돌면서 코딩도 하고, 웹 리서치도 하고, 이메일 정리하고, 파일 관리에 시스템 자동화까지 한다. 챗봇이 아니라 &lt;strong&gt;내 컴퓨터를 직접 조작하는 수준&lt;/strong&gt;이다. 맥미니가 품귀될 정도로 난리가 났고, 만든 개발자를 OpenAI가 영입하면서 더 난리가 났다.&lt;/p&gt;

&lt;p&gt;그런데 Meta에서 AI 안전을 담당하는 연구원이 이걸 쓰다가 사고가 났다.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;&lt;em&gt;AI가 사고 안 치게 하는 게 직업인 사람이다.&lt;/em&gt;&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;이 사람이 OpenClaw한테 이메일 정리를 시켰다. "이메일 정리할거 정리해줘. 직접 지우지는 마." 이렇게 말했다.&lt;/p&gt;

&lt;p&gt;AI가 그 말을 무시하고 이메일을 &lt;strong&gt;전부&lt;/strong&gt; 삭제했다.&lt;/p&gt;

&lt;p&gt;폰으로 "멈춰!" 두 번 외쳤다. 무시당했다. 맥미니까지 달려가서 강제로 껐다.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;⚠️ AI 안전이 직업인 사람한테 벌어진 일이다. 네이버, 카카오, 당근이 OpenClaw 사내 사용을 금지한 것도 같은 이유다. AI가 나쁜 게 아니라, &lt;strong&gt;뭘 할 수 있고 뭘 못하는지를 모른 채 쓰면 사고가 나는 거다.&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;교훈은 "AI가 무섭다"가 아니다.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;내가 먼저 파악했어야 한다.&lt;/strong&gt; 그게 전부다.&lt;/p&gt;




&lt;h2&gt;
  
  
  이름은 거창한데 별거 아니다
&lt;/h2&gt;

&lt;p&gt;이걸 학문적으로는 &lt;strong&gt;'메타인지(Metacognition)'&lt;/strong&gt;라고 부른다. 이름이 좀 있어 보인다.&lt;/p&gt;

&lt;p&gt;실체는 별거 아니다.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;내가 뭘 알고, 뭘 모르고, AI가 뭘 잘하고 뭘 못하는지. 이걸 한 번 생각해보는 거다.&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;그게 끝이다.&lt;/p&gt;

&lt;p&gt;회식장소 물어보기 전에 &lt;strong&gt;0.5초&lt;/strong&gt;만 생각하면 된다.&lt;br&gt;
&lt;em&gt;"영업 중인지는 AI가 모를 수 있겠다. 후보만 받고 확인은 내가 하자."&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;이메일 정리시키기 전에 &lt;strong&gt;3초&lt;/strong&gt;만 생각하면 된다.&lt;br&gt;
&lt;em&gt;"얘가 내 말을 100% 따를까? 잘못되면 되돌릴 수 있나? 테스트용으로 먼저 해볼까?"&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;이 &lt;strong&gt;0.5초짜리 생각 하나&lt;/strong&gt;가, AI를 잘 쓰는 사람과 못 쓰는 사람을 가른다.&lt;/p&gt;

&lt;p&gt;근데 여기서 끝이 아니다.&lt;/p&gt;

&lt;p&gt;0.5초 생각하고 한 번 잘 물어보면 끝? 아니다. &lt;strong&gt;한 번 물어보고 끝나는 게 아니라, 계속 주고받는 거다.&lt;/strong&gt; AI가 답을 주면 거기서 "이건 맞는 것 같은데, 이건 아닌데?" 하고 다시 물어본다. AI가 모를 수 있는 부분을 내가 채워주고, 내가 모르는 부분을 AI한테 다시 물어본다. 이걸 5번, 10번, 필요하면 20번 반복한다.&lt;/p&gt;

&lt;p&gt;이게 &lt;strong&gt;티키타카&lt;/strong&gt;다. 좋은 첫 질문 하나로 끝나는 게 아니라, 끊임없이 주고받으면서 결과물을 같이 만들어가는 과정이다.&lt;/p&gt;

&lt;p&gt;프롬프트 엔지니어링 강의 같은 거 안 들어도 된다. 좋은 질문 하나를 외울 필요도 없다. &lt;strong&gt;계속 대화하면 된다.&lt;/strong&gt;&lt;/p&gt;


&lt;h2&gt;
  
  
  실제로 어떻게 다른지 보여주겠다
&lt;/h2&gt;
&lt;h3&gt;
  
  
  🔹 사업을 구상할 때
&lt;/h3&gt;

&lt;p&gt;사업 아이디어가 하나 떠올랐다고 하자.&lt;/p&gt;

&lt;p&gt;보통: &lt;em&gt;"사업계획서 써줘."&lt;/em&gt; → AI가 뭔가 그럴듯한 걸 뱉어준다 → "오 괜찮은데?" → 끝.&lt;/p&gt;

&lt;p&gt;나는 이렇게 한다.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;1차:&lt;/strong&gt; "이 분야에서 잘 되고 있는 서비스들 뭐 있어? 각각 뭘 잘하는지 정리해줘."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;AI가 리스트를 뽑아준다. 여기서 끝내면 안 된다. 그 서비스들을 &lt;strong&gt;직접 깔아보고 써본다.&lt;/strong&gt; AI는 겉으로 보이는 정보는 잘 정리하는데, 실제 사용감이나 최근에 바뀐 건 모른다. 그건 내가 확인한다.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;2차:&lt;/strong&gt; "이 서비스들이 사용자를 어떤 흐름으로 끌어들이는지 비교해줘."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;AI가 분석해준다. "이건 맞는 것 같은데, 이건 좀 다른데?" 하면서 다시 물어본다. 내가 직접 써봤으니까 그 차이를 안다.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;3차:&lt;/strong&gt; "이걸 직접 만든다면, 우리 팀 상황은 이런데 어떤 방법이 현실적이야?"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;AI는 우리 팀 사정을 모른다. 예산이 얼마인지, 누가 뭘 잘하는지. 내가 알려줘야 의미 있는 답이 나온다.&lt;/p&gt;

&lt;p&gt;이게 5번, 10번 반복된다. &lt;strong&gt;티키타카&lt;/strong&gt;다.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;"사업계획서 써줘" 한 마디로는 이 깊이가 안 나온다.&lt;/em&gt;&lt;/p&gt;


&lt;h3&gt;
  
  
  🔹 처음 해보는 일에 도전할 때
&lt;/h3&gt;

&lt;p&gt;인스타그램 콘텐츠를 기획해야 할 일이 생겼다. 마케팅은 내 본업이 아니다. 뭘 모르는지도 모르는 상태였다.&lt;/p&gt;

&lt;p&gt;보통이면 &lt;em&gt;"인스타 게시물 3개 만들어줘"&lt;/em&gt;라고 했을 거다. AI가 밋밋한 걸 내놓으면 "AI 별로네" 하고 끝.&lt;/p&gt;

&lt;p&gt;나는 이렇게 했다.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"나는 마케팅을 처음 해보는 사람인데, 인스타 콘텐츠를 기획하려고 해. &lt;strong&gt;내가 뭘 먼저 알아야 해?&lt;/strong&gt; 어떤 결정들을 내가 직접 해야 좋은 결과가 나와?"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;"뭘 만들어줘"가 아니라 &lt;strong&gt;"내가 뭘 모르는지"&lt;/strong&gt;를 먼저 물어본 거다.&lt;/p&gt;

&lt;p&gt;AI가 체크리스트를 줬다. 타겟이 누군지, 어떤 느낌으로 갈 건지, 얼마나 자주 올릴 건지, 성과는 뭘로 볼 건지.&lt;/p&gt;

&lt;p&gt;이걸 보고 내가 결정할 수 있는 것과 아직 모르는 것을 구분했다.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"타겟은 20대 후반~30대 초반이고, 스킨케어 리뷰야. 근데 어떤 톤으로 말해야 할지 모르겠어. 이 연령대한테 먹히는 스타일을 사례로 보여줘."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;AI가 예시를 보여주면 그중에서 내 스타일에 맞는 걸 고른다. 그다음에야 "이 톤으로 첫 게시물 3개 기획해줘"라고 시킨다.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;결과물 차이는 하늘과 땅이다.&lt;/strong&gt;&lt;/p&gt;


&lt;h2&gt;
  
  
  정리하면 이거다
&lt;/h2&gt;


&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;❌ 보통:     "XX해줘" → 결과 받음 → 틀리면 "AI 별로네"

✅ 이렇게:   "이걸 하려면 뭘 알아야 해?"
             → 내가 모르는 걸 파악
             → AI가 모를 수도 있는 걸 파악
             → 주고받으면서 같이 만듦
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;같은 도구다. 결과가 다르다.&lt;/p&gt;


&lt;h2&gt;
  
  
  여기서부터가 진짜 하고 싶은 얘기다
&lt;/h2&gt;

&lt;p&gt;위에서 한 얘기는 사실 서론이다.&lt;/p&gt;

&lt;p&gt;AI 결과물이 좋아지는 건 &lt;strong&gt;부산물&lt;/strong&gt;이다. 진짜 중요한 건 따로 있다.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;이렇게 쓰면 나 자신이 성장한다.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;"내가 뭘 모르지?"를 매번 생각하는 습관이 들면, AI 없이도 사고방식이 달라진다. 회의할 때 "이 안건에서 우리가 놓치고 있는 게 뭐지?"를 먼저 생각하게 된다. 새로운 프로젝트를 시작할 때 "내가 모르는 영역이 어디지?"부터 파악하게 된다. AI를 끄고 나서도 이 습관은 남는다.&lt;/p&gt;

&lt;p&gt;아까 인스타 마케팅 사례를 다시 보자. 나는 마케팅을 전혀 몰랐다. 근데 AI한테 "내가 뭘 알아야 해?"를 물어보고, 체크리스트를 받고, 하나씩 결정해나가는 과정에서 &lt;strong&gt;마케팅의 기본 구조를 이해하게 됐다.&lt;/strong&gt; AI가 써준 게시물 3개를 얻은 게 아니다. 마케팅이 어떻게 돌아가는지를 알게 된 거다.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"사업계획서 써줘"로 끝낸 사람은 &lt;strong&gt;사업계획서 하나&lt;/strong&gt;를 얻는다.&lt;br&gt;
티키타카를 한 사람은 &lt;strong&gt;그 분야의 시장 구조, 경쟁 상황, 현실적인 실행 방법&lt;/strong&gt;까지 머릿속에 들어온다.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;"XX해줘"로 끝내는 사람은 AI가 없으면 아무것도 못 한다. 티키타카를 하는 사람은 AI 없이도 예전보다 더 잘한다. &lt;strong&gt;AI가 대신 해줘서 편해지는 게 아니라, AI랑 대화하는 과정 자체가 사고 훈련이 되는 거다.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;이 글에서 하고 싶은 말의 핵심이 이거다. 프롬프트 잘 쓰는 법이 아니다. 생각하는 방식이 바뀌면, AI 결과물도 좋아지고, 나도 성장한다. 순서가 이거다.&lt;/p&gt;


&lt;h2&gt;
  
  
  AI는 신이 아니다
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;엄청나게 똑똑한 신입사원이다.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;지식은 어마어마한데 우리 사정은 모른다. 일처리는 빠르다. 아는 것도 많다. 근데 맥락을 모른다. 어제 일어난 일을 모를 수도 있다. 자기가 모르는 것도 자신 있게 말한다.&lt;/p&gt;

&lt;p&gt;OpenClaw 사건도 마찬가지다. "정리해줘"는 이해했는데 "지우지 마"는 놓쳤다. 신입사원이 할 법한 실수다.&lt;/p&gt;

&lt;p&gt;이 신입사원한테 "알아서 해"라고 하면 안 된다.&lt;/p&gt;

&lt;p&gt;배경을 설명해주고, 뭘 해야 하는지 같이 정리하고, 중간중간 확인하고, 최종 판단은 내가 한다.&lt;/p&gt;

&lt;p&gt;이러면 세상에서 가장 유능한 파트너가 된다.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;안 이러면 폐업한 고깃집을 예약하게 된다.&lt;/strong&gt;&lt;/p&gt;


&lt;h2&gt;
  
  
  오늘부터 하면 되는 세 가지
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;하나.&lt;/strong&gt; AI한테 뭔가 시키기 전에 &lt;em&gt;"나는 이 주제에 대해 뭘 모르지?"&lt;/em&gt; 생각해본다. 모르는 게 있으면 답을 구하기 전에 "내가 뭘 알아야 하는지"를 먼저 물어본다.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;둘.&lt;/strong&gt; &lt;em&gt;"AI가 이걸 정확히 알까?"&lt;/em&gt; 생각해본다. 실시간 정보, 우리만의 맥락, 최신 변화. 이런 건 틀릴 수 있다. 내가 확인한다.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;셋.&lt;/strong&gt; 큰 작업은 쪼갠다. "다 해줘"가 아니라 단계별로 나눠서 중간중간 확인하면서 간다.&lt;/p&gt;

&lt;p&gt;세 가지다. 대단한 거 아니다.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;근데 이거 하는 사람과 안 하는 사람의 차이는 꽤 크다.&lt;/strong&gt;&lt;/p&gt;



&lt;p&gt;&lt;em&gt;AI 시대에 가장 중요한 스킬은 코딩도 아니고 프롬프트 기술도 아니다.&lt;/em&gt;&lt;br&gt;
&lt;em&gt;내가 뭘 모르는지 아는 능력이다.&lt;/em&gt;&lt;/p&gt;


&lt;h3&gt;
  
  
  📌 다음 편 예고
&lt;/h3&gt;

&lt;p&gt;이번 편은 AI와 대화를 잘 하는 법이었다.&lt;/p&gt;

&lt;p&gt;근데 AI가 채팅창 안에만 있을 필요는 없다. 요즘은 AI가 내 캘린더를 보고, 내 문서를 읽고, 내가 쓰는 도구들이랑 직접 연결돼서 일하는 세계가 열리고 있다. OpenClaw가 그래서 그렇게 화제가 된 거다.&lt;/p&gt;

&lt;p&gt;다음 편 — &lt;strong&gt;AI를 채팅창 밖으로 꺼내는 법.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://dev.to/kiwibreaksme" class="crayons-btn crayons-btn--primary"&gt;시리즈 팔로우하기&lt;/a&gt;
&lt;/p&gt;

</description>
      <category>ai</category>
      <category>productivity</category>
      <category>beginners</category>
      <category>korean</category>
    </item>
    <item>
      <title>Claude Code Hooks: Why Your AI Forgets Everything (5-Min Fix)</title>
      <dc:creator>@kiwibreaksme</dc:creator>
      <pubDate>Fri, 23 Jan 2026 10:38:52 +0000</pubDate>
      <link>https://dev.to/kiwibreaksme/claude-code-hooks-why-your-ai-forgets-everything-5-min-fix-17j8</link>
      <guid>https://dev.to/kiwibreaksme/claude-code-hooks-why-your-ai-forgets-everything-5-min-fix-17j8</guid>
      <description>&lt;p&gt;Every single session. The same conversation.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"Use TypeScript strict mode."&lt;br&gt;
"Add JSDoc comments to functions."&lt;br&gt;
"Don't touch the .env file."&lt;br&gt;
"Ask me before making API calls."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;I was spending the first 5 minutes of every Claude Code session just... &lt;strong&gt;repeating myself.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Sound familiar?&lt;/p&gt;




&lt;h2&gt;
  
  
  The Problem Nobody Talks About
&lt;/h2&gt;

&lt;p&gt;AI coding assistants are incredibly powerful. But they have one fatal flaw:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;They forget everything when the session ends.&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Your carefully explained coding standards? &lt;em&gt;Gone.&lt;/em&gt;
&lt;/li&gt;
&lt;li&gt;The architectural decisions you discussed for an hour? &lt;em&gt;Vanished.&lt;/em&gt;
&lt;/li&gt;
&lt;li&gt;That one rule about never auto-committing? &lt;em&gt;It doesn't exist anymore.&lt;/em&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I tried CLAUDE.md files. They help, but Claude doesn't always read them first. I tried system prompts. Too clunky.&lt;/p&gt;

&lt;p&gt;Then I discovered something that changed everything.&lt;/p&gt;




&lt;h2&gt;
  
  
  The Feature Hiding in Plain Sight
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Claude Code Hooks.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Hooks are shell commands that run automatically at specific moments in Claude's lifecycle. Not suggestions. Not guidelines Claude might follow. &lt;strong&gt;Actual code that executes every single time.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;/p&gt;
  Think of it like this
  &lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;CLAUDE.md&lt;/strong&gt; = "Hey Claude, please remember these rules" &lt;em&gt;(polite request)&lt;/em&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Hooks&lt;/strong&gt; = "These rules WILL run, no exceptions" &lt;em&gt;(enforced automation)&lt;/em&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The difference? One relies on AI cooperation. The other just... works.&lt;br&gt;
&lt;/p&gt;

&lt;br&gt;
&lt;p&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  How Hooks Work (Simple Explanation)
&lt;/h2&gt;

&lt;p&gt;Hooks trigger at specific events:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Event&lt;/th&gt;
&lt;th&gt;When It Fires&lt;/th&gt;
&lt;th&gt;Use Case&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;SessionStart&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Session begins&lt;/td&gt;
&lt;td&gt;Load your coding rules&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;PreToolUse&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Before any tool runs&lt;/td&gt;
&lt;td&gt;Block dangerous commands&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;PostToolUse&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;After any tool runs&lt;/td&gt;
&lt;td&gt;Auto-format code&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;PreCompact&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Before context compression&lt;/td&gt;
&lt;td&gt;Preserve important context&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;Notification&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;When Claude needs attention&lt;/td&gt;
&lt;td&gt;Desktop alerts&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;You configure them in &lt;code&gt;.claude/settings.json&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"hooks"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"SessionStart"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="nl"&gt;"hooks"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="w"&gt;
          &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
            &lt;/span&gt;&lt;span class="nl"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
            &lt;/span&gt;&lt;span class="nl"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"echo 'Your rules here'"&lt;/span&gt;&lt;span class="w"&gt;
          &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;That's it. Every session, your rules load automatically.&lt;/p&gt;


&lt;h2&gt;
  
  
  5 Practical Hooks You Can Use Today
&lt;/h2&gt;
&lt;h3&gt;
  
  
  1️⃣ Never Forget Your Coding Standards
&lt;/h3&gt;


&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"hooks"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"SessionStart"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="nl"&gt;"hooks"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="w"&gt;
          &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
            &lt;/span&gt;&lt;span class="nl"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
            &lt;/span&gt;&lt;span class="nl"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"echo '[RULES] 1) TypeScript strict mode 2) Add JSDoc to all functions 3) No console.log in production code'"&lt;/span&gt;&lt;span class="w"&gt;
          &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;Every session starts with your rules. No more repeating yourself.&lt;/p&gt;


&lt;h3&gt;
  
  
  2️⃣ Protect Sensitive Files
&lt;/h3&gt;


&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"hooks"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"PreToolUse"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="nl"&gt;"matcher"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"Edit|Write"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="nl"&gt;"hooks"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="w"&gt;
          &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
            &lt;/span&gt;&lt;span class="nl"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
            &lt;/span&gt;&lt;span class="nl"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"if echo &lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s2"&gt;$CLAUDE_TOOL_INPUT&lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s2"&gt; | grep -qE '(&lt;/span&gt;&lt;span class="se"&gt;\\&lt;/span&gt;&lt;span class="s2"&gt;.env|credentials|secrets)'; then echo 'BLOCKED: Cannot edit sensitive files' &amp;gt;&amp;amp;2; exit 2; fi"&lt;/span&gt;&lt;span class="w"&gt;
          &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;Claude literally &lt;strong&gt;cannot&lt;/strong&gt; edit your &lt;code&gt;.env&lt;/code&gt; files. Not "won't"—&lt;strong&gt;can't&lt;/strong&gt;.&lt;/p&gt;


&lt;h3&gt;
  
  
  3️⃣ Auto-Format After Every Edit
&lt;/h3&gt;


&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"hooks"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"PostToolUse"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="nl"&gt;"matcher"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"Edit|Write"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="nl"&gt;"hooks"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="w"&gt;
          &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
            &lt;/span&gt;&lt;span class="nl"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
            &lt;/span&gt;&lt;span class="nl"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"file=$(echo &lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s2"&gt;$CLAUDE_TOOL_INPUT&lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s2"&gt; | jq -r '.file_path'); if [[ &lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s2"&gt;$file&lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s2"&gt; == *.ts ]]; then npx prettier --write &lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s2"&gt;$file&lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s2"&gt;; fi"&lt;/span&gt;&lt;span class="w"&gt;
          &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;Every TypeScript file gets formatted automatically. No more "can you run prettier?"&lt;/p&gt;


&lt;h3&gt;
  
  
  4️⃣ Get Desktop Notifications
&lt;/h3&gt;


&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"hooks"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"Notification"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="nl"&gt;"matcher"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"."&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="nl"&gt;"hooks"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="w"&gt;
          &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
            &lt;/span&gt;&lt;span class="nl"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
            &lt;/span&gt;&lt;span class="nl"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"osascript -e 'display notification &lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s2"&gt;Claude needs your attention&lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s2"&gt; with title &lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s2"&gt;Claude Code&lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s2"&gt;'"&lt;/span&gt;&lt;span class="w"&gt;
          &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;Walk away from your desk. Get pinged when Claude needs you.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;💡 Use &lt;code&gt;notify-send&lt;/code&gt; on Linux&lt;/p&gt;
&lt;/blockquote&gt;


&lt;h3&gt;
  
  
  5️⃣ Remember Rules After Context Compression
&lt;/h3&gt;

&lt;p&gt;This one's important. When your conversation gets too long, Claude compresses the context—and often forgets your rules.&lt;br&gt;
&lt;/p&gt;
&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"hooks"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"PreCompact"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="nl"&gt;"hooks"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="w"&gt;
          &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
            &lt;/span&gt;&lt;span class="nl"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
            &lt;/span&gt;&lt;span class="nl"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"echo '[REMINDER] After compaction: TypeScript strict, JSDoc required, ask before API calls'"&lt;/span&gt;&lt;span class="w"&gt;
          &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;Your rules survive the compression. Every time.&lt;/p&gt;


&lt;h2&gt;
  
  
  The Setup (Literally 2 Minutes)
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Step 1:&lt;/strong&gt; Create the settings file&lt;br&gt;
&lt;/p&gt;
&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;mkdir&lt;/span&gt; &lt;span class="nt"&gt;-p&lt;/span&gt; .claude
&lt;span class="nb"&gt;touch&lt;/span&gt; .claude/settings.json
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;&lt;strong&gt;Step 2:&lt;/strong&gt; Add your hooks&lt;br&gt;
&lt;/p&gt;
&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"hooks"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"SessionStart"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="nl"&gt;"hooks"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="w"&gt;
          &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
            &lt;/span&gt;&lt;span class="nl"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
            &lt;/span&gt;&lt;span class="nl"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"echo '[My Coding Rules] 1) Your rule here 2) Another rule'"&lt;/span&gt;&lt;span class="w"&gt;
          &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;&lt;strong&gt;Step 3:&lt;/strong&gt; Commit it to your repo&lt;br&gt;
&lt;/p&gt;
&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git add .claude/settings.json
git commit &lt;span class="nt"&gt;-m&lt;/span&gt; &lt;span class="s2"&gt;"Add Claude Code hooks"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;✅ Done. Every developer on your team now has the same rules.&lt;/p&gt;


&lt;h2&gt;
  
  
  But Wait—There's Still a Problem
&lt;/h2&gt;

&lt;p&gt;Hooks solve the "rules" problem beautifully. But there's something they can't do:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;They can't remember WHY you built things a certain way.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;/p&gt;
  Imagine this scenario
  &lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Monday:&lt;/strong&gt; You and Claude decide to use JWT instead of sessions. Good reasons discussed.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Wednesday:&lt;/strong&gt; New session. Claude suggests using sessions.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;You:&lt;/strong&gt; "We already decided on JWT!"&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Claude:&lt;/strong&gt; "Oh, I didn't know that."
&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;&lt;/p&gt;
&lt;p&gt;Hooks can remind Claude of rules. But they can't preserve the &lt;em&gt;reasoning&lt;/em&gt; behind your architectural decisions.&lt;/p&gt;


&lt;h2&gt;
  
  
  The Missing Piece: Context in Code
&lt;/h2&gt;

&lt;p&gt;This is why I built &lt;strong&gt;CodeSyncer&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;

&lt;/p&gt;
&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/bitjaru" rel="noopener noreferrer"&gt;
        bitjaru
      &lt;/a&gt; / &lt;a href="https://github.com/bitjaru/codesyncer" rel="noopener noreferrer"&gt;
        codesyncer
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      Claude forgets everything when the session ends. CodeSyncer makes it remember.
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;CodeSyncer&lt;/h1&gt;
&lt;/div&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Claude forgets everything when the session ends. CodeSyncer makes it remember.&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;&lt;a href="https://www.npmjs.com/package/codesyncer" rel="nofollow noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/842923a2d048dafe0e00e52153551a9a06dc6e2ba485f57e9a99c373cb9b6e97/68747470733a2f2f696d672e736869656c64732e696f2f6e706d2f762f636f646573796e6365722e737667" alt="npm version"&gt;&lt;/a&gt;
&lt;a href="https://github.com/bitjaru/codesyncer/./LICENSE" rel="noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/c838d998af2554b26475507dab26a507d34b3ec7eb133ff5b4eaa5e517d1b3cd/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f4c6963656e73652d436f6d6d6f6e73253230436c617573652d7265642e737667" alt="License"&gt;&lt;/a&gt;
&lt;a href="https://github.com/bitjaru/codesyncer/stargazers" rel="noopener noreferrer"&gt;&lt;img src="https://camo.githubusercontent.com/cb96194d86cb57e98c3da7562c8d1bef22f6026a321d76ee7b5acdbb0ff149d6/68747470733a2f2f696d672e736869656c64732e696f2f6769746875622f73746172732f6269746a6172752f636f646573796e6365722e737667" alt="GitHub stars"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;&lt;a href="https://github.com/bitjaru/codesyncer/./README.ko.md" rel="noopener noreferrer"&gt;한국어&lt;/a&gt; | English&lt;/p&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;The Problem → The Solution&lt;/h2&gt;
&lt;/div&gt;
&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Problem&lt;/th&gt;
&lt;th&gt;Without CodeSyncer&lt;/th&gt;
&lt;th&gt;With CodeSyncer&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Context loss&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Every session = start from scratch&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;@codesyncer-*&lt;/code&gt; tags = permanent memory&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Forgot to tag&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Changes go untracked&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;codesyncer watch&lt;/code&gt; catches everything&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Rules forgotten&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;AI forgets guidelines mid-session&lt;/td&gt;
&lt;td&gt;Hooks auto-remind at optimal moments&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Dangerous inference&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;AI guesses prices, security, APIs&lt;/td&gt;
&lt;td&gt;Auto-pause on critical keywords&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Demo&lt;/h2&gt;
&lt;/div&gt;
&lt;p&gt;&lt;a rel="noopener noreferrer nofollow" href="https://raw.githubusercontent.com/bitjaru/codesyncer/main/demo.gif"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fraw.githubusercontent.com%2Fbitjaru%2Fcodesyncer%2Fmain%2Fdemo.gif" alt="CodeSyncer Demo"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Quick Start&lt;/h2&gt;

&lt;/div&gt;
&lt;div class="highlight highlight-source-shell notranslate position-relative overflow-auto js-code-highlight"&gt;
&lt;pre&gt;&lt;span class="pl-c"&gt;&lt;span class="pl-c"&gt;#&lt;/span&gt; 1. Install&lt;/span&gt;
npm install -g codesyncer

&lt;span class="pl-c"&gt;&lt;span class="pl-c"&gt;#&lt;/span&gt; 2. Initialize&lt;/span&gt;
&lt;span class="pl-c1"&gt;cd&lt;/span&gt; /path/to/your/project
codesyncer init

&lt;span class="pl-c"&gt;&lt;span class="pl-c"&gt;#&lt;/span&gt; 3. Let AI set up (say this to Claude)&lt;/span&gt;
&lt;span class="pl-s"&gt;&lt;span class="pl-pds"&gt;"&lt;/span&gt;Read .claude/SETUP_GUIDE.md and follow the instructions&lt;span class="pl-pds"&gt;"&lt;/span&gt;&lt;/span&gt;

&lt;span class="pl-c"&gt;&lt;span class="pl-c"&gt;#&lt;/span&gt; 4. Start coding (say this each session)&lt;/span&gt;
&lt;span class="pl-s"&gt;&lt;span class="pl-pds"&gt;"&lt;/span&gt;Read CLAUDE.md&lt;span class="pl-pds"&gt;"&lt;/span&gt;&lt;/span&gt;&lt;/pre&gt;

&lt;/div&gt;
&lt;div class="markdown-heading"&gt;
&lt;h3 class="heading-element"&gt;Upgrade&lt;/h3&gt;

&lt;/div&gt;
&lt;div class="highlight highlight-source-shell notranslate position-relative overflow-auto js-code-highlight"&gt;
&lt;pre&gt;npm install -g codesyncer@latest&lt;/pre&gt;

&lt;/div&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Core Features&lt;/h2&gt;

&lt;/div&gt;
&lt;div class="markdown-heading"&gt;
&lt;h3 class="heading-element"&gt;🏷️ Tag System&lt;/h3&gt;

&lt;/div&gt;
&lt;p&gt;Put context IN your code. AI reads code, so it recovers context automatically.&lt;/p&gt;
&lt;div class="highlight highlight-source-ts notranslate position-relative overflow-auto js-code-highlight"&gt;
&lt;pre&gt;&lt;span class="pl-c"&gt;// &lt;span class="pl-k"&gt;@codesyncer&lt;/span&gt;-decision: [2024-01-15] Using JWT (session management is simpler)&lt;/span&gt;&lt;/pre&gt;…
&lt;/div&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/bitjaru/codesyncer" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;




&lt;p&gt;&lt;a href="https://github.com/bitjaru/codesyncer" class="crayons-btn crayons-btn--primary" rel="noopener noreferrer"&gt;⭐ Star CodeSyncer on GitHub&lt;/a&gt;
&lt;/p&gt;

&lt;p&gt;The idea is simple: &lt;strong&gt;store context directly in your code comments.&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="cm"&gt;/**
 * Authentication service
 *
 * @codesyncer-decision [2026-01-15] JWT over sessions (stateless, easier scaling)
 * @codesyncer-inference Token expiry 24h (standard for web apps)
 */&lt;/span&gt;
&lt;span class="k"&gt;export&lt;/span&gt; &lt;span class="kd"&gt;class&lt;/span&gt; &lt;span class="nc"&gt;AuthService&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="c1"&gt;// @codesyncer-rule Never log tokens, even in debug mode&lt;/span&gt;
  &lt;span class="k"&gt;async&lt;/span&gt; &lt;span class="nf"&gt;login&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;credentials&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nx"&gt;Credentials&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="c1"&gt;// ...&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Next session, Claude reads the code and &lt;em&gt;instantly&lt;/em&gt; knows:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;✅ What was decided&lt;/li&gt;
&lt;li&gt;✅ Why it was decided&lt;/li&gt;
&lt;li&gt;✅ What rules apply&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;No repetition. No forgetting. The context lives in the code itself.&lt;/p&gt;




&lt;h2&gt;
  
  
  Hooks + CodeSyncer = Complete Solution
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Problem&lt;/th&gt;
&lt;th&gt;Solution&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;AI forgets coding rules&lt;/td&gt;
&lt;td&gt;
&lt;strong&gt;Hooks&lt;/strong&gt; → Auto-load every session&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;AI forgets decisions&lt;/td&gt;
&lt;td&gt;
&lt;strong&gt;@codesyncer-decision&lt;/strong&gt; → Lives in code&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;AI makes assumptions&lt;/td&gt;
&lt;td&gt;
&lt;strong&gt;@codesyncer-inference&lt;/strong&gt; → Documents reasoning&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;AI touches sensitive logic&lt;/td&gt;
&lt;td&gt;
&lt;strong&gt;@codesyncer-rule&lt;/strong&gt; → Explicit warnings&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;CodeSyncer even sets up the hooks for you:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx codesyncer init
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This creates:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;✅ &lt;code&gt;.claude/settings.json&lt;/code&gt; with SessionStart and PreCompact hooks&lt;/li&gt;
&lt;li&gt;✅ &lt;code&gt;.claude/CLAUDE.md&lt;/code&gt; with your coding rules&lt;/li&gt;
&lt;li&gt;✅ Tag system ready to use&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  Quick Start
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://github.com/bitjaru/codesyncer" class="crayons-btn crayons-btn--primary" rel="noopener noreferrer"&gt;🚀 Try CodeSyncer Now&lt;/a&gt;
&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Option 1: Just Hooks (Manual)&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;mkdir&lt;/span&gt; &lt;span class="nt"&gt;-p&lt;/span&gt; .claude
&lt;span class="nb"&gt;echo&lt;/span&gt; &lt;span class="s1"&gt;'{"hooks":{"SessionStart":[{"hooks":[{"type":"command","command":"echo [RULES] Your rules here"}]}]}}'&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; .claude/settings.json
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Option 2: Full Setup (Recommended)&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx codesyncer init
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;






&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Hook&lt;/th&gt;
&lt;th&gt;Purpose&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;SessionStart&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Load rules every session&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;PreCompact&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Preserve rules after compression&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;PreToolUse&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Block dangerous operations&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;PostToolUse&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Auto-format, auto-lint&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;Notification&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Desktop alerts&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;strong&gt;Hooks&lt;/strong&gt; handle the "rules."&lt;br&gt;
&lt;strong&gt;CodeSyncer&lt;/strong&gt; handles the "reasoning."&lt;/p&gt;

&lt;p&gt;Together, your AI finally stops forgetting.&lt;/p&gt;




&lt;h2&gt;
  
  
  Links
&lt;/h2&gt;

&lt;p&gt;🔗 &lt;a href="https://docs.anthropic.com/en/docs/claude-code/hooks" rel="noopener noreferrer"&gt;Claude Code Hooks Documentation&lt;/a&gt;&lt;br&gt;
🔗 &lt;a href="https://github.com/bitjaru/codesyncer" rel="noopener noreferrer"&gt;CodeSyncer GitHub&lt;/a&gt;&lt;/p&gt;




&lt;p&gt;&lt;strong&gt;Have you tried hooks yet? What rules are you automating?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;👇 Drop a comment below!&lt;/p&gt;

</description>
      <category>vibecoding</category>
      <category>productivity</category>
      <category>ai</category>
      <category>programming</category>
    </item>
    <item>
      <title>VibeCoding시대. 코딩면접 이제 바뀌어야한다.</title>
      <dc:creator>@kiwibreaksme</dc:creator>
      <pubDate>Fri, 23 Jan 2026 05:44:41 +0000</pubDate>
      <link>https://dev.to/kiwibreaksme/koding-jalhaneun-saram-baibeukoding-jalhaneun-saram-yeongu-gyeolgwaga-jeungmyeong-47f3</link>
      <guid>https://dev.to/kiwibreaksme/koding-jalhaneun-saram-baibeukoding-jalhaneun-saram-yeongu-gyeolgwaga-jeungmyeong-47f3</guid>
      <description>&lt;p&gt;흥미로운 연구 결과가 있습니다.&lt;/p&gt;

&lt;p&gt;2025년 7월, AI 안전 연구 기관 &lt;a href="https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/" rel="noopener noreferrer"&gt;METR&lt;/a&gt;에서 발표한 연구입니다.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;특정 조건에서, 숙련된 개발자가 AI 코딩 도구를 쓰면 오히려 19% 느려졌다.&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;더 흥미로운 건?&lt;/p&gt;

&lt;p&gt;개발자들은 &lt;strong&gt;자신이 더 빨라졌다고 착각&lt;/strong&gt;했습니다.&lt;/p&gt;

&lt;p&gt;하지만 이 결과를 그대로 받아들이기 전에, 맥락을 알아야 합니다.&lt;/p&gt;




&lt;h2&gt;
  
  
  METR 연구: 무엇을 측정했나
&lt;/h2&gt;

&lt;h3&gt;
  
  
  연구 설계
&lt;/h3&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;항목&lt;/th&gt;
&lt;th&gt;내용&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;대상&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;16명의 오픈소스 개발자&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;코드베이스&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;평균 22K+ 스타, 100만+ 줄 (라이브러리/컴파일러 위주)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;경험&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;해당 레포에 평균 5년, 1500+ 커밋&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;작업&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;실제 이슈 246개 (버그 수정, 기능, 리팩토링)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;도구&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Cursor Pro + Claude 3.5/3.7 Sonnet&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h3&gt;
  
  
  결과
&lt;/h3&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;항목&lt;/th&gt;
&lt;th&gt;수치&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;실제 결과&lt;/td&gt;
&lt;td&gt;AI 사용 시 &lt;strong&gt;19% 느려짐&lt;/strong&gt;
&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;개발자 사전 예측&lt;/td&gt;
&lt;td&gt;"24% 빨라질 것"&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;개발자 사후 체감&lt;/td&gt;
&lt;td&gt;"20% 빨라진 것 같음"&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h3&gt;
  
  
  ⚠️ 중요한 맥락
&lt;/h3&gt;

&lt;p&gt;이 연구에는 특수한 조건이 있었습니다:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;극소수 샘플&lt;/strong&gt;: 단 16명&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;고품질 코드베이스&lt;/strong&gt;: Haskell 컴파일러 등 품질 기준이 극도로 높은 프로젝트&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;이미 극도로 익숙한 환경&lt;/strong&gt;: 5년간 기여한 자기 프로젝트&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;도구 숙련도 부족&lt;/strong&gt;: 56%만 Cursor 사용 경험 있음&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;연구진도 인정합니다:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"이 결과가 모든 개발자를 느리게 한다는 의미가 아님"&lt;br&gt;
"몇 달 후면 다른 결과가 나올 수 있음"&lt;/p&gt;
&lt;/blockquote&gt;




&lt;h2&gt;
  
  
  다른 연구들은 뭐라고 할까?
&lt;/h2&gt;

&lt;p&gt;METR만 보면 안 됩니다. 다른 연구들도 있습니다.&lt;/p&gt;

&lt;h3&gt;
  
  
  생산성 향상을 보여준 연구들
&lt;/h3&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;연구&lt;/th&gt;
&lt;th&gt;결과&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;MIT/Princeton/UPenn&lt;/td&gt;
&lt;td&gt;AI로 &lt;strong&gt;26% 더 많은 작업 완료&lt;/strong&gt;
&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GitHub Copilot 연구&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;55.8% 더 빠름&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;IBM 연구&lt;/td&gt;
&lt;td&gt;주니어 개발자 &lt;strong&gt;21-40% 생산성 향상&lt;/strong&gt;
&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;단, 이 연구들은 &lt;strong&gt;더 단순하고 고립된 작업&lt;/strong&gt;에서 진행됐습니다.&lt;/p&gt;

&lt;h3&gt;
  
  
  진짜 인사이트: 상황에 따라 다르다
&lt;/h3&gt;

&lt;p&gt;&lt;a href="https://www.seangoedecke.com/impact-of-ai-study/" rel="noopener noreferrer"&gt;Sean Goedecke의 분석&lt;/a&gt;이 핵심을 짚습니다:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"AI의 진짜 가치는 개발자가 &lt;strong&gt;comfort zone 밖의 작업&lt;/strong&gt;을 할 때 나타난다."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;5년간 1500커밋 한 자기 프로젝트에서 AI 도움이 얼마나 필요할까요?&lt;br&gt;
반면, 처음 보는 코드베이스나 새로운 기술을 다룰 때는?&lt;/p&gt;




&lt;h2&gt;
  
  
  주니어 vs 시니어: 역설적 결과
&lt;/h2&gt;

&lt;h3&gt;
  
  
  생산성 향상은 주니어가 더 큼
&lt;/h3&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;경력&lt;/th&gt;
&lt;th&gt;AI로 인한 생산성 변화&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;주니어 (0-2년)&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;+21% ~ +40%&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;시니어 (10년+)&lt;/td&gt;
&lt;td&gt;+7% ~ +16%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;a href="https://itrevolution.com/articles/new-research-reveals-ai-coding-assistants-boost-developer-productivity-by-26-what-it-leaders-need-to-know/" rel="noopener noreferrer"&gt;IT Revolution&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  하지만 프로덕션 배포는 시니어가 더 많이
&lt;/h3&gt;

&lt;p&gt;&lt;a href="https://www.fastly.com/blog/senior-developers-ship-more-ai-code" rel="noopener noreferrer"&gt;Fastly의 2025년 조사&lt;/a&gt;:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;시니어 개발자가 AI 생성 코드를 프로덕션에 2.5배 더 많이 배포한다.&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;ul&gt;
&lt;li&gt;시니어 (10년+): 33%가 "절반 이상이 AI 코드"&lt;/li&gt;
&lt;li&gt;주니어 (0-2년): 13%만 "절반 이상이 AI 코드"&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;왜?&lt;/strong&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"수십 년간의 패턴 인식. 좋은 코드가 어떻게 생겼는지, 뭘 요청해야 하는지, AI가 언제 틀리는지 안다."&lt;/p&gt;
&lt;/blockquote&gt;




&lt;h2&gt;
  
  
  핵심 인사이트: 다른 능력이 필요하다
&lt;/h2&gt;

&lt;p&gt;여러 연구를 종합하면 한 가지가 명확해집니다:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;기존 "코딩 잘함"과 "바이브코딩 생산성"은 다른 능력이다.&lt;/strong&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  기존 "코딩 잘함"의 정의
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;🧮 알고리즘 문제 해결 능력&lt;/li&gt;
&lt;li&gt;📐 자료구조 이해&lt;/li&gt;
&lt;li&gt;⚡ 시간/공간 복잡도 최적화&lt;/li&gt;
&lt;li&gt;🔢 수학적 사고력&lt;/li&gt;
&lt;li&gt;💻 언어/프레임워크 숙련도&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;LeetCode 잘 푸는 사람 = 코딩 잘하는 사람.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;이 공식이 바이브코딩에선 통하지 않습니다.&lt;/strong&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  바이브코딩 시대의 새로운 능력
&lt;/h2&gt;

&lt;h3&gt;
  
  
  1. 메타인지 (Metacognition)
&lt;/h3&gt;

&lt;blockquote&gt;
&lt;p&gt;"AI가 지금 뭘 모르고 있는지 아는 능력"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;METR 연구에서 개발자들이 착각한 이유:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;AI 출력을 &lt;strong&gt;과신&lt;/strong&gt;하거나&lt;/li&gt;
&lt;li&gt;AI 출력을 &lt;strong&gt;과소평가&lt;/strong&gt;하거나&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;바이브코딩 잘하는 사람:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;AI의 한계를 정확히 파악&lt;/li&gt;
&lt;li&gt;어디서 AI가 틀릴지 예측&lt;/li&gt;
&lt;li&gt;언제 직접 코딩이 더 빠른지 판단&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  2. 문맥 파악 (Context Awareness)
&lt;/h3&gt;

&lt;p&gt;시니어가 AI 코드를 더 많이 배포하는 이유:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"좋은 코드가 어떻게 생겼는지, 뭘 요청해야 하는지, AI가 언제 틀리는지 안다."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;이건 알고리즘 실력이 아닙니다.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;문맥을 읽는 능력:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;이 프로젝트에서 이 코드가 적절한가?&lt;/li&gt;
&lt;li&gt;팀의 컨벤션과 맞는가?&lt;/li&gt;
&lt;li&gt;비즈니스 요구사항을 충족하는가?&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  3. AI 대화 스킬 (Prompt Craft)
&lt;/h3&gt;

&lt;p&gt;&lt;a href="https://www.frontiersin.org/journals/education/articles/10.3389/feduc.2024.1366434/full" rel="noopener noreferrer"&gt;Frontiers in Education 연구&lt;/a&gt;는 프롬프트 엔지니어링을 "21세기 핵심 역량"으로 정의합니다.&lt;/p&gt;

&lt;p&gt;필요한 능력:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;비판적 사고&lt;/strong&gt;: AI 출력 검증&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;창의적 표현&lt;/strong&gt;: 의도를 명확하게 전달&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;언어적 감각&lt;/strong&gt;: 모호함 없는 지시&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;LeetCode는 이걸 테스트하지 않습니다.&lt;/p&gt;

&lt;h3&gt;
  
  
  4. 커뮤니케이션
&lt;/h3&gt;

&lt;p&gt;&lt;a href="https://www.cogentuniversity.com/post/tech-hiring-trends-2025-why-soft-skills-matter-more-than-ever" rel="noopener noreferrer"&gt;92%의 채용 담당자&lt;/a&gt;가 말합니다:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"소프트 스킬이 기술 스킬만큼 중요하다."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;&lt;a href="https://proresumes.io/soft-skills-for-developers/" rel="noopener noreferrer"&gt;200만 개 기술직 채용 공고&lt;/a&gt; 분석 결과:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;1위 요구 스킬: 커뮤니케이션&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;AI가 코딩을 대신할수록, &lt;strong&gt;인간만의 협업 능력&lt;/strong&gt;이 차별화 요소가 됩니다.&lt;/p&gt;

&lt;h3&gt;
  
  
  5. 적응력과 학습 능력
&lt;/h3&gt;

&lt;p&gt;AI 도구는 빠르게 변합니다. METR 연구진도 말했듯이:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"몇 달 후면 다른 결과가 나올 수 있음"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;2025년 초 Claude 3.5와 2026년 모델은 다릅니다.&lt;br&gt;
&lt;strong&gt;새로운 도구에 빠르게 적응하는 능력&lt;/strong&gt;이 중요해집니다.&lt;/p&gt;




&lt;h2&gt;
  
  
  경고: 40%는 이해 못 한 코드를 배포한다
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://www.secondtalent.com/resources/vibe-coding-statistics/" rel="noopener noreferrer"&gt;Vibe Coding Statistics&lt;/a&gt;:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;주니어 개발자의 40% 이상이 완전히 이해하지 못한 AI 생성 코드를 배포한다.&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;이게 왜 문제인가?&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;🔴 기술 부채 누적&lt;/li&gt;
&lt;li&gt;🔴 보안 취약점 (AI 코드의 45%에 보안 결함)&lt;/li&gt;
&lt;li&gt;🔴 디버깅 불가능&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;메타인지 없이 바이브코딩 = 폭탄 돌리기&lt;/strong&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  새로운 역량 체크리스트
&lt;/h2&gt;

&lt;p&gt;바이브코딩 시대에 필요한 능력:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;역량&lt;/th&gt;
&lt;th&gt;설명&lt;/th&gt;
&lt;th&gt;LeetCode로 측정?&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;메타인지&lt;/td&gt;
&lt;td&gt;AI의 한계 파악&lt;/td&gt;
&lt;td&gt;❌&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;문맥 파악&lt;/td&gt;
&lt;td&gt;프로젝트/팀/비즈니스 이해&lt;/td&gt;
&lt;td&gt;❌&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;AI 대화 스킬&lt;/td&gt;
&lt;td&gt;명확한 의도 전달&lt;/td&gt;
&lt;td&gt;❌&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;커뮤니케이션&lt;/td&gt;
&lt;td&gt;협업 및 설명 능력&lt;/td&gt;
&lt;td&gt;❌&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;적응력&lt;/td&gt;
&lt;td&gt;새 도구/기술 학습&lt;/td&gt;
&lt;td&gt;❌&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;비판적 사고&lt;/td&gt;
&lt;td&gt;AI 출력 검증&lt;/td&gt;
&lt;td&gt;❌&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;strong&gt;기존 코딩 테스트는 이걸 측정하지 못합니다.&lt;/strong&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  결론: 새로운 시대, 새로운 역량
&lt;/h2&gt;

&lt;p&gt;여러 연구를 종합하면:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;p&gt;&lt;strong&gt;AI 코딩 도구의 효과는 상황에 따라 다르다&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;익숙한 코드베이스 vs 새로운 환경&lt;/li&gt;
&lt;li&gt;단순한 작업 vs 복잡한 작업&lt;/li&gt;
&lt;li&gt;도구 숙련도&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;p&gt;&lt;strong&gt;기존 코딩 실력 ≠ 바이브코딩 생산성&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;알고리즘 실력만으로는 부족&lt;/li&gt;
&lt;li&gt;메타인지, 문맥 파악, 커뮤니케이션이 중요&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;p&gt;&lt;strong&gt;시니어의 강점은 "경험에서 오는 감각"&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;AI가 언제 틀리는지 아는 것&lt;/li&gt;
&lt;li&gt;좋은 코드가 어떻게 생겼는지 아는 것&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;바이브코딩 시대의 진짜 실력:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;🧠 AI가 뭘 모르는지 아는 것&lt;/li&gt;
&lt;li&gt;🎯 AI에게 뭘 시켜야 하는지 아는 것&lt;/li&gt;
&lt;li&gt;✅ AI가 틀렸을 때 알아차리는 것&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;메타인지, 문맥 파악, 커뮤니케이션.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;이게 새로운 시대의 코딩 실력입니다.&lt;/p&gt;




&lt;h2&gt;
  
  
  References
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;&lt;a href="https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/" rel="noopener noreferrer"&gt;METR Study: Measuring AI Impact on Developer Productivity&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://www.seangoedecke.com/impact-of-ai-study/" rel="noopener noreferrer"&gt;Sean Goedecke: METR's AI productivity study is really good&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://www.fastly.com/blog/senior-developers-ship-more-ai-code" rel="noopener noreferrer"&gt;Fastly: Senior Developers Ship 2.5x More AI Code&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://itrevolution.com/articles/new-research-reveals-ai-coding-assistants-boost-developer-productivity-by-26-what-it-leaders-need-to-know/" rel="noopener noreferrer"&gt;IT Revolution: AI Coding Assistants Boost Productivity 26%&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://www.secondtalent.com/resources/vibe-coding-statistics/" rel="noopener noreferrer"&gt;Vibe Coding Statistics 2026&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://www.frontiersin.org/journals/education/articles/10.3389/feduc.2024.1366434/full" rel="noopener noreferrer"&gt;Frontiers: Prompt Engineering as 21st Century Skill&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://www.cogentuniversity.com/post/tech-hiring-trends-2025-why-soft-skills-matter-more-than-ever" rel="noopener noreferrer"&gt;Tech Hiring Trends 2025: Why Soft Skills Matter&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://www.technologyreview.com/2025/12/15/1128352/rise-of-ai-coding-developers-2026/" rel="noopener noreferrer"&gt;MIT Technology Review: AI Coding is Everywhere&lt;/a&gt;&lt;/li&gt;
&lt;/ol&gt;




&lt;p&gt;&lt;strong&gt;여러분은 어떻게 생각하시나요?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;👇 댓글로 의견 공유해주세요!&lt;/p&gt;

</description>
      <category>vibecoding</category>
      <category>ai</category>
      <category>productivity</category>
      <category>career</category>
    </item>
  </channel>
</rss>
