<?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: Yash</title>
    <description>The latest articles on DEV Community by Yash (@yashh1323).</description>
    <link>https://dev.to/yashh1323</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%2F4254951%2F90a6db5f-9cea-44e4-9e7d-d2c31a4305f0.jpg</url>
      <title>DEV Community: Yash</title>
      <link>https://dev.to/yashh1323</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/yashh1323"/>
    <language>en</language>
    <item>
      <title>Why AI Essay Detectors Fail, So I Built a Rust Proof-of-Write Engine</title>
      <dc:creator>Yash</dc:creator>
      <pubDate>Fri, 20 Nov 2026 10:50:59 +0000</pubDate>
      <link>https://dev.to/yashh1323/why-ai-essay-detectors-fail-so-i-built-a-rust-proof-of-write-engine-4fej</link>
      <guid>https://dev.to/yashh1323/why-ai-essay-detectors-fail-so-i-built-a-rust-proof-of-write-engine-4fej</guid>
      <description>&lt;p&gt;Post-hoc AI essay detectors fail in both directions at once. Non-native English writers who use formal grammar get falsely flagged by perplexity classifiers all the time, while passing generated prose through a basic paraphrase prompt walks right past those same checks. When an accusation lands, the teacher is staring at a single opaque number and the student can't show they spent three hours actually drafting and revising.&lt;/p&gt;

&lt;p&gt;That setup is unfair to honest writers and irritatingly easy to game. Worse, the usual edtech fix of browser lockdowns and continuous webcam recording crosses privacy lines I wanted nothing to do with.&lt;/p&gt;

&lt;p&gt;So I built &lt;a href="https://github.com/yashh1323/Clio" rel="noopener noreferrer"&gt;&lt;code&gt;Clio&lt;/code&gt;&lt;/a&gt; (MIT), an open-source writing studio and authorship auditor that records &lt;strong&gt;how&lt;/strong&gt; an essay was composed inside the browser canvas rather than guessing from finished text, backed by a Pure-Rust verification engine (&lt;code&gt;engine-rust/&lt;/code&gt;). Live demo: &lt;a href="https://clio-three.vercel.app" rel="noopener noreferrer"&gt;clio-three.vercel.app&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fffalj7q16a51z13qta6x.gif" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fffalj7q16a51z13qta6x.gif" alt="Clio Working Session Replay and Authorship Intelligence Demo" width="800" height="380"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Two design rules: no spyware, no black-box scores
&lt;/h2&gt;

&lt;p&gt;Before touching any detection math, I wrote down two hard invariants in the codebase.&lt;/p&gt;

&lt;p&gt;First, Clio won't use webcams, microphones, or OS-level hooks. It runs inside a standard Next.js 16 + TipTap ProseMirror editor with &lt;code&gt;IndexedDB&lt;/code&gt; (&lt;code&gt;ClioDB&lt;/code&gt;) offline autosave and Supabase PostgreSQL (&lt;code&gt;7&lt;/code&gt; tables protected by Row Level Security), observing only what happens inside the editor canvas.&lt;/p&gt;

&lt;p&gt;Second, there are no opaque percentages on their own. Instead of outputting "84% AI," Clio hands educators a scrubbable &lt;strong&gt;Session Replay&lt;/strong&gt; timeline alongside physical signals inside the document (discrete paste events and keystroke cadence histograms), so a student can actually point to their editing history.&lt;/p&gt;

&lt;p&gt;On the Rust side (&lt;code&gt;engine-rust&lt;/code&gt;), the compute service enforces &lt;code&gt;#![forbid(unsafe_code)]&lt;/code&gt;, makes zero outbound network calls, and compares internal engine API keys in constant time using bitwise XOR folding. If Rust isn't running locally (for example, on a serverless preview deployment), a Strangler Fig proxy in Next.js (&lt;code&gt;lib/engine.ts&lt;/code&gt;) falls back cleanly to TypeScript.&lt;/p&gt;

&lt;h2&gt;
  
  
  Four independent verification layers in Pure Rust (&lt;code&gt;Axum 0.8&lt;/code&gt; + &lt;code&gt;Rayon&lt;/code&gt;)
&lt;/h2&gt;

&lt;p&gt;Any single detector has an obvious bypass. Instead of betting on one metric, &lt;code&gt;engine-rust&lt;/code&gt; evaluates four separate layers.&lt;/p&gt;

&lt;h3&gt;
  
  
  1. Lexical Winnowing &amp;amp; Containment (&lt;code&gt;src/winnowing.rs&lt;/code&gt;)
&lt;/h3&gt;

&lt;p&gt;Raw substring hashing misses edited patches, while comparing global Jaccard similarity misses a copied paragraph buried inside a long original essay. Clio implements Schleimer, Wilkerson, and Aiken's Winnowing algorithm (&lt;code&gt;k = 5&lt;/code&gt;, guarantee threshold &lt;code&gt;t = 12&lt;/code&gt;, sliding window &lt;code&gt;w = t - k + 1 = 8&lt;/code&gt;) over 64-bit FNV-1a shingle hashes with deterministic rightmost tie-breaking:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight rust"&gt;&lt;code&gt;&lt;span class="k"&gt;pub&lt;/span&gt; &lt;span class="k"&gt;const&lt;/span&gt; &lt;span class="no"&gt;NOISE_THRESHOLD_K&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;usize&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;5&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;pub&lt;/span&gt; &lt;span class="k"&gt;const&lt;/span&gt; &lt;span class="no"&gt;GUARANTEE_THRESHOLD_T&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;usize&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;12&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;pub&lt;/span&gt; &lt;span class="k"&gt;const&lt;/span&gt; &lt;span class="no"&gt;WINDOW_SIZE_W&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;usize&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="no"&gt;GUARANTEE_THRESHOLD_T&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="no"&gt;NOISE_THRESHOLD_K&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;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Instead of symmetric Jaccard, it computes asymmetric Broder containment &lt;code&gt;C(A, B) = |S(A) ∩ S(B)| / |S(A)|&lt;/code&gt;, so lifting 200 words out of a 2,000-word source still flags a 90% collision inside that block without blowing up the index.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. Parallel MinHash LSH (&lt;code&gt;src/similarity.rs&lt;/code&gt;)
&lt;/h3&gt;

&lt;p&gt;To compare an essay against a whole class cohort without &lt;code&gt;O(N^2)&lt;/code&gt; full-text scans, &lt;code&gt;similarity.rs&lt;/code&gt; computes 128 universal hash permutations in parallel using &lt;code&gt;rayon::prelude::*&lt;/code&gt; and buckets candidates with 64x2 Locality-Sensitive Hashing bands.&lt;/p&gt;

&lt;h3&gt;
  
  
  3. Intrinsic Stylometry &amp;amp; Semantic Projection (&lt;code&gt;src/stylometry.rs&lt;/code&gt;, &lt;code&gt;src/semantic.rs&lt;/code&gt;)
&lt;/h3&gt;

&lt;p&gt;If a student writes half an essay in their normal voice and pastes generated or ghostwritten paragraphs into the middle, global checks miss the seam. &lt;code&gt;stylometry.rs&lt;/code&gt; slides a 250-word window (&lt;code&gt;stride = 50&lt;/code&gt;) tracking 50 English function words via Burrows' Delta alongside Yule's length-invariant Characteristic &lt;code&gt;K&lt;/code&gt;. Alongside it, &lt;code&gt;semantic.rs&lt;/code&gt; projects words and character trigrams into a pure-Rust 64-dimensional vector space with synonym clustering to catch word-swapped paraphrasing.&lt;/p&gt;

&lt;h3&gt;
  
  
  4. Log-Normal Keystroke Process Telemetry (&lt;code&gt;src/telemetry_advanced.rs&lt;/code&gt;)
&lt;/h3&gt;

&lt;p&gt;Human typing roughly follows Log-Normal Inter-Keystroke Interval (IKI) distributions punctured by planning pauses (&lt;code&gt;P-bursts&lt;/code&gt;) and backspacing cycles (&lt;code&gt;R-bursts&lt;/code&gt;). Macro scripts and typing bots either emit unnaturally uniform intervals or fake randomness with flat uniform jitter.&lt;/p&gt;

&lt;p&gt;To avoid pulling C bindings or external crates just to evaluate a normal CDF in the Kolmogorov-Smirnov test (&lt;code&gt;D_KS&lt;/code&gt;), &lt;code&gt;telemetry_advanced.rs&lt;/code&gt; implements a zero-allocation &lt;code&gt;erf_approx()&lt;/code&gt; using Abramowitz &amp;amp; Stegun Formula 7.1.26 (&lt;code&gt;&amp;lt; 1.5e-7&lt;/code&gt; error):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight rust"&gt;&lt;code&gt;&lt;span class="k"&gt;pub&lt;/span&gt; &lt;span class="k"&gt;struct&lt;/span&gt; &lt;span class="nc"&gt;ProcessAuditResult&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="k"&gt;pub&lt;/span&gt; &lt;span class="n"&gt;is_authentic&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;bool&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="k"&gt;pub&lt;/span&gt; &lt;span class="n"&gt;confidence_score&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;f64&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="k"&gt;pub&lt;/span&gt; &lt;span class="n"&gt;p_burst_mean&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;f64&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="k"&gt;pub&lt;/span&gt; &lt;span class="n"&gt;r_burst_mean&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;f64&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="k"&gt;pub&lt;/span&gt; &lt;span class="n"&gt;revision_ratio&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;f64&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="k"&gt;pub&lt;/span&gt; &lt;span class="n"&gt;cognitive_pause_ratio&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;f64&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="k"&gt;pub&lt;/span&gt; &lt;span class="n"&gt;log_mean&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;f64&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="k"&gt;pub&lt;/span&gt; &lt;span class="n"&gt;log_std_dev&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;f64&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="k"&gt;pub&lt;/span&gt; &lt;span class="n"&gt;ks_d_statistic&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;f64&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="k"&gt;pub&lt;/span&gt; &lt;span class="n"&gt;rapid_interval_count&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;usize&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="k"&gt;pub&lt;/span&gt; &lt;span class="n"&gt;valid_interval_count&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;usize&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="k"&gt;pub&lt;/span&gt; &lt;span class="n"&gt;flags&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;Vec&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;&lt;/span&gt;&lt;span class="kt"&gt;String&lt;/span&gt;&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="k"&gt;pub&lt;/span&gt; &lt;span class="n"&gt;method&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;String&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;Export reports don't spawn headless Chrome or Node PDF wrappers either: &lt;code&gt;src/pdf.rs&lt;/code&gt; writes raw ISO 32000-1 PDF 1.4 byte offsets directly in Rust.&lt;/p&gt;

&lt;h2&gt;
  
  
  Verification &amp;amp; live demo accounts
&lt;/h2&gt;

&lt;p&gt;Right now &lt;code&gt;engine-rust&lt;/code&gt; passes &lt;code&gt;69/69&lt;/code&gt; Cargo tests (&lt;code&gt;cargo clippy --all-targets -- -D warnings&lt;/code&gt; clean) and the web tier passes &lt;code&gt;31/31&lt;/code&gt; security and E2E Playwright QA suites (&lt;code&gt;auto_qa_tests/&lt;/code&gt;).&lt;/p&gt;

&lt;p&gt;You can click through the live app without registering first at &lt;strong&gt;&lt;a href="https://clio-three.vercel.app/login" rel="noopener noreferrer"&gt;clio-three.vercel.app/login&lt;/a&gt;&lt;/strong&gt;:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Teacher login (&lt;code&gt;teacher&lt;/code&gt; / &lt;code&gt;ClioTeacher2026!&lt;/code&gt;) for Session Replay and similarity reports&lt;/li&gt;
&lt;li&gt;Student login (&lt;code&gt;student&lt;/code&gt; / &lt;code&gt;ClioStudent2026!&lt;/code&gt;) for the TipTap writing studio (faculty signup invite code: &lt;code&gt;CLIO-FACULTY-2026&lt;/code&gt;)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Repo (MIT): &lt;a href="https://github.com/yashh1323/Clio" rel="noopener noreferrer"&gt;github.com/yashh1323/Clio&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What browser software still can't stop
&lt;/h2&gt;

&lt;p&gt;There is one boundary Clio deliberately doesn't cross. If someone puts an AI-generated draft on a second monitor and manually types every sentence into the TipTap editor at human speed (pausing occasionally, fixing a typo here and there), no purely web-sandboxed telemetry will catch it reliably. Stopping that requires cameras, keyloggers, or locked-down hardware, which breaks trust with every honest student in the room.&lt;/p&gt;

</description>
      <category>showdev</category>
      <category>rust</category>
      <category>nextjs</category>
      <category>opensource</category>
    </item>
    <item>
      <title>Frontend Skills Kept Breaking My Dashboards, So I Added Chrome CDP</title>
      <dc:creator>Yash</dc:creator>
      <pubDate>Fri, 20 Nov 2026 09:27:00 +0000</pubDate>
      <link>https://dev.to/yashh1323/frontend-skills-kept-breaking-my-dashboards-so-i-added-chrome-cdp-353b</link>
      <guid>https://dev.to/yashh1323/frontend-skills-kept-breaking-my-dashboards-so-i-added-chrome-cdp-353b</guid>
      <description>&lt;p&gt;I've been using frontend skills like Anthropic's &lt;code&gt;frontend-design&lt;/code&gt; and &lt;code&gt;ui-ux-pro-max&lt;/code&gt; in Claude Code and Cursor. On blank landing pages they help a lot. Things kept going sideways whenever I dropped them into an existing React or Next.js codebase just to fix a settings view or clean up a data table.&lt;/p&gt;

&lt;p&gt;Half the time the agent put oversized editorial headings into an admin layout, or wrapped sections in &lt;code&gt;100vw&lt;/code&gt; and gave mobile a horizontal scrollbar. And once a long session hit &lt;code&gt;/compact&lt;/code&gt;, whatever layout rules I'd typed at the start fell out of context anyway.&lt;/p&gt;

&lt;p&gt;I wanted the agent to stop guessing purely from prompts and verify what Chrome actually renders, so I split my setup into an MIT skill called &lt;a href="https://github.com/yashh1323/atelier-ui" rel="noopener noreferrer"&gt;&lt;code&gt;atelier-ui&lt;/code&gt;&lt;/a&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.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fqx0s5y2ynre5zx6m90zs.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%2Fqx0s5y2ynre5zx6m90zs.png" alt="atelier-ui verified builds across Brand Sites, SaaS Consoles, and 3D WebGL Shaders" width="800" height="400"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Splitting one giant prompt into a ~5 KB router
&lt;/h2&gt;

&lt;p&gt;Keeping 1,500 lines of marketing typography rules loaded while you're editing an admin table doesn't end well.&lt;/p&gt;

&lt;p&gt;Instead of dumping everything upfront, &lt;a href="https://github.com/yashh1323/atelier-ui/blob/main/SKILL.md" rel="noopener noreferrer"&gt;&lt;code&gt;SKILL.md&lt;/code&gt;&lt;/a&gt; acts as a tiny router with three steps:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;code&gt;scripts/detect-env.mjs&lt;/code&gt; reads &lt;code&gt;package.json&lt;/code&gt; and checks which local browsers are installed once, so the skill never assumes global CLIs are present.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;01-intake.md&lt;/code&gt; settles open questions in one turn and writes the brief to &lt;code&gt;.atelier/BRIEF.md&lt;/code&gt;. Even if a long session compacts twice, the constraints stay on disk.&lt;/li&gt;
&lt;li&gt;The agent loads only the build module for your task: &lt;code&gt;02-brand-site.md&lt;/code&gt; for marketing pages, &lt;code&gt;03-app-ui.md&lt;/code&gt; for dashboards and existing codebases, or &lt;code&gt;04-3d-visual.md&lt;/code&gt; for Three.js and WebGL scenes.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The line in &lt;code&gt;03-app-ui.md&lt;/code&gt; that mattered most in practice is disabling landing-page hero gating and dealt themes from &lt;code&gt;library/themes.json&lt;/code&gt; completely. Existing Tailwind tokens and &lt;code&gt;components/ui/&lt;/code&gt; primitives stay untouched, numeric tables keep &lt;code&gt;tabular-nums&lt;/code&gt;, and controls have to handle default, hover/focus-visible, disabled, loading, empty, and error states.&lt;/p&gt;

&lt;h2&gt;
  
  
  Checking mechanical CSS bugs at 0 LLM tokens
&lt;/h2&gt;

&lt;p&gt;Asking an agent in prose to remember &lt;code&gt;prefers-reduced-motion&lt;/code&gt; and avoid &lt;code&gt;transition: all&lt;/code&gt; works for roughly 200 lines. Moving those checks into &lt;a href="https://github.com/yashh1323/atelier-ui/blob/main/scripts/lint-ui.mjs" rel="noopener noreferrer"&gt;&lt;code&gt;scripts/lint-ui.mjs&lt;/code&gt;&lt;/a&gt; takes them out of the prompt completely.&lt;/p&gt;

&lt;p&gt;Here's one rule from &lt;code&gt;lint-ui.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="p"&gt;{&lt;/span&gt;
  &lt;span class="n"&gt;id&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'no-100vw'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;severity&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'FAIL'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;exts&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s1"&gt;'.css'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s1"&gt;'.html'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s1"&gt;'.jsx'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s1"&gt;'.tsx'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s1"&gt;'.vue'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s1"&gt;'.svelte'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s1"&gt;'.astro'&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;test&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;line&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt;\&lt;span class="n"&gt;b100vw&lt;/span&gt;\&lt;span class="n"&gt;b&lt;/span&gt;&lt;span class="o"&gt;|&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="o"&gt;^&lt;/span&gt;\&lt;span class="n"&gt;w&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="n"&gt;w&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nf"&gt;screen&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="n"&gt;w&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="o"&gt;/&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;test&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;line&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;msg&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'Avoid `100vw` / `w-screen` (causes horizontal scrollbar overflow). Use `100%` / `w-full`.'&lt;/span&gt;&lt;span class="p"&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;Two small fixes cut out most of the false positives when running this on real repos. &lt;code&gt;stripCommentsPreserveLines()&lt;/code&gt; replaces &lt;code&gt;&amp;lt;!-- --&amp;gt;&lt;/code&gt;, &lt;code&gt;{/* */}&lt;/code&gt;, and &lt;code&gt;/* */&lt;/code&gt; comment blocks with spaces while keeping line breaks intact (and ignores &lt;code&gt;https://&lt;/code&gt; when stripping &lt;code&gt;//&lt;/code&gt; comments) so reported line numbers still match your editor. The &lt;code&gt;no-outline-none&lt;/code&gt; rule also explicitly allows Tailwind patterns like &lt;code&gt;focus-visible:outline-none focus-visible:ring-2&lt;/code&gt; instead of rejecting valid focus rings.&lt;/p&gt;

&lt;h2&gt;
  
  
  Checking real Chrome pixels without installing Playwright
&lt;/h2&gt;

&lt;p&gt;Even clean CSS breaks on mobile once real webfonts and sticky headers render. At the same time, forcing everyone who installs a &lt;code&gt;SKILL.md&lt;/code&gt; to pull a 200 MB Playwright and Chromium bundle slows down quick setups and clutters &lt;code&gt;package.json&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;05-qa-gate.md&lt;/code&gt; delegates first: if &lt;code&gt;impeccable&lt;/code&gt; or &lt;code&gt;agent-browser&lt;/code&gt; is already on your &lt;code&gt;PATH&lt;/code&gt;, it uses that. When neither is installed (the usual case on a clean machine), &lt;a href="https://github.com/yashh1323/atelier-ui/blob/main/scripts/check-ui.mjs" rel="noopener noreferrer"&gt;&lt;code&gt;scripts/check-ui.mjs&lt;/code&gt;&lt;/a&gt; runs with zero npm dependencies by driving whichever Chrome, Edge, Chromium, or Brave binary is already on disk over Node 22's built-in &lt;code&gt;WebSocket&lt;/code&gt; client.&lt;/p&gt;

&lt;p&gt;In one pass it starts a local static server (&lt;code&gt;resolveSafeFile()&lt;/code&gt; blocks &lt;code&gt;..&lt;/code&gt; and sibling-prefix path escapes) or attaches to &lt;code&gt;--url http://localhost:3000&lt;/code&gt;. It fails immediately on the layout regressions that hurt most in practice: &lt;code&gt;1440x900&lt;/code&gt; or &lt;code&gt;390x844&lt;/code&gt; horizontal overflow, and sticky headers overlapping &lt;code&gt;&amp;lt;h1&amp;gt;&lt;/code&gt;. The same run also checks for unloaded webfonts, zero-size SVGs or canvases, &lt;code&gt;oklch&lt;/code&gt; WCAG AA contrast errors, &lt;code&gt;console.error&lt;/code&gt; output, and blank WebGL frames under &lt;code&gt;--shots&lt;/code&gt;. Passing &lt;code&gt;--full --sheet&lt;/code&gt; stitches desktop and mobile captures into one &lt;code&gt;review-sheet.png&lt;/code&gt;, so a zero-context critic pass only needs a single image turn.&lt;/p&gt;

&lt;h2&gt;
  
  
  3 live builds and how to try it
&lt;/h2&gt;

&lt;p&gt;I put together a three-mode showcase (&lt;a href="https://yashh1323.github.io/atelier-ui-showcase/" rel="noopener noreferrer"&gt;yashh1323.github.io/atelier-ui-showcase&lt;/a&gt;) with Nordhavn Timber (brand site), DriftGate Console (dark ops UI), and Meridian 04 (WebGL2 astrolabe). Eleven headless Chrome smoke tests in &lt;code&gt;tests/smoke.mjs&lt;/code&gt; run against them in GitHub Actions (&lt;code&gt;docs/PROOFS.md&lt;/code&gt;).&lt;/p&gt;

&lt;p&gt;Include the GitHub prefix if you test the installer, because bare &lt;code&gt;atelier-ui&lt;/code&gt; on npm belongs to a different author:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx github:yashh1323/atelier-ui install
npx github:yashh1323/atelier-ui doctor

&lt;span class="c"&gt;# Or via vercel-labs/skills&lt;/span&gt;
npx skills add yashh1323/atelier-ui
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Repo: &lt;a href="https://github.com/yashh1323/atelier-ui" rel="noopener noreferrer"&gt;github.com/yashh1323/atelier-ui&lt;/a&gt; (MIT)&lt;/p&gt;

&lt;p&gt;It won't rescue an unclear brief or supply good design taste on its own, and &lt;code&gt;check-ui.mjs&lt;/code&gt; needs Node 22+ plus a Chromium-family browser installed on the machine.&lt;/p&gt;

</description>
      <category>showdev</category>
      <category>webdev</category>
      <category>ai</category>
      <category>opensource</category>
    </item>
  </channel>
</rss>
