<?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: neo xia</title>
    <description>The latest articles on DEV Community by neo xia (@neo_xia_3f4c019330af5fb9d).</description>
    <link>https://dev.to/neo_xia_3f4c019330af5fb9d</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%2F3961904%2F1a660dc3-e1fb-4caf-b9fe-c28d09fd5a72.png</url>
      <title>DEV Community: neo xia</title>
      <link>https://dev.to/neo_xia_3f4c019330af5fb9d</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/neo_xia_3f4c019330af5fb9d"/>
    <language>en</language>
    <item>
      <title>HEX to Pantone is a guess: making color conversion honest in an interface</title>
      <dc:creator>neo xia</dc:creator>
      <pubDate>Mon, 16 Nov 2026 02:36:45 +0000</pubDate>
      <link>https://dev.to/neo_xia_3f4c019330af5fb9d/hex-to-pantone-is-a-guess-making-color-conversion-honest-in-an-interface-3930</link>
      <guid>https://dev.to/neo_xia_3f4c019330af5fb9d/hex-to-pantone-is-a-guess-making-color-conversion-honest-in-an-interface-3930</guid>
      <description>&lt;p&gt;Every color tool eventually ships the same sentence: &lt;em&gt;closest Pantone match&lt;/em&gt;. It sounds authoritative, and almost every implementation behind it is an approximation. That gap between the confident label and the actual math is worth designing around, because someone will print a business card from whatever your UI claims.&lt;/p&gt;

&lt;h2&gt;
  
  
  Three different conversions are hiding under one button
&lt;/h2&gt;

&lt;p&gt;A HEX value is an sRGB triplet with a defined transfer curve. RGB to CMYK conversion is device dependent: the real answer depends on the press, the ink set, the substrate and the ICC profile used to separate the file. Without a profile you have a formula, not a measurement. A CMYK-to-Pantone lookup is a different approximation again, because Pantone references are ink recipes and physical swatches, not RGB values.&lt;/p&gt;

&lt;p&gt;So a converter that says "this CMYK is that Pantone" is guessing at two levels. That is fine as long as the interface says so.&lt;/p&gt;

&lt;h2&gt;
  
  
  Distance is the part most tools skip
&lt;/h2&gt;

&lt;p&gt;If you map a color to a reference library, you have to rank candidates by something. Euclidean distance in raw RGB is convenient and visibly wrong: two colors that differ by the same RGB amount do not look equally different. Converting to CIELAB first and measuring with CIEDE2000 gets much closer to perceived difference, which matters most in the greens and blues where sRGB is coarsest.&lt;/p&gt;

&lt;p&gt;Here is the shape of the pipeline in plain terms:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Parse the input into sRGB (or take CMYK percentages as given).&lt;/li&gt;
&lt;li&gt;Convert sRGB to XYZ under a D65 white point.&lt;/li&gt;
&lt;li&gt;Convert XYZ to CIELAB.&lt;/li&gt;
&lt;li&gt;Measure CIEDE2000 against every candidate reference.&lt;/li&gt;
&lt;li&gt;Return the nearest match together with the distance, not just the name.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;That last step is the honest one. A nearest neighbor with a distance of 1.2 and one with a distance of 8 are not the same answer, and the user deserves to know which they got.&lt;/p&gt;

&lt;h2&gt;
  
  
  Design rules that keep the UI truthful
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Show the distance.&lt;/strong&gt; "Nearest reference, ΔE 3.4" is information. "Exact match" is a lie unless it is one.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Separate coated and uncoated.&lt;/strong&gt; They are different swatch books and they produce different nearest references. Collapsing them into one answer creates a wrong answer for half the users.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Keep the input reproducible.&lt;/strong&gt; If a row in a chart shows a match, link to the converter state that produced it, so the number can be checked instead of trusted.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;State the limits in the interface, not only in a footnote.&lt;/strong&gt; CMYK without a profile, digital approximations of a licensed library, and the need for a physical proof all belong next to the result.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What this looks like in practice
&lt;/h2&gt;

&lt;p&gt;I built &lt;a href="https://hueconvert.com" rel="noopener noreferrer"&gt;HueConvert&lt;/a&gt; around those rules: a free converter that takes HEX, RGB, CMYK or HSL, returns the other formats, and shows the nearest Pantone coated and uncoated reference with the CIEDE2000 distance attached. The three tool pages carry printable charts of 139 references each, and every row links back to the exact input that produced it. No account, no upload, and the charts render without JavaScript.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Disclosure: I operate HueConvert. It is free, has no paid placement, and this post is a write-up of the conversion decisions, not a launch announcement.&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;Color conversion is a chain of approximations. The engineering work is not pretending otherwise; it is choosing the least misleading approximation, exposing the distance, and making the result reproducible. Users forgive an approximate answer that is labeled as approximate. They do not forgive a confident one that ruins a print run.&lt;/p&gt;

</description>
      <category>webdev</category>
      <category>css</category>
      <category>design</category>
      <category>color</category>
    </item>
    <item>
      <title>How I verify an ONNX model swap by hashing tensors (and what it still doesn't prove)</title>
      <dc:creator>neo xia</dc:creator>
      <pubDate>Thu, 12 Nov 2026 16:23:36 +0000</pubDate>
      <link>https://dev.to/neo_xia_3f4c019330af5fb9d/how-i-verify-an-onnx-model-swap-by-hashing-tensors-and-what-it-still-doesnt-prove-15a3</link>
      <guid>https://dev.to/neo_xia_3f4c019330af5fb9d/how-i-verify-an-onnx-model-swap-by-hashing-tensors-and-what-it-still-doesnt-prove-15a3</guid>
      <description>&lt;p&gt;We run background removal as an ONNX model inside a container behind a Cloudflare Worker. Last week we split one combined runtime image into two profile images — a "Lite" profile and a "Lite-2K" profile — and shrunk the distributed weights. The interesting part was not the image split. It was proving that the split changed nothing about the output.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why "the model still loads" proves nothing
&lt;/h2&gt;

&lt;p&gt;ONNX models are frequently stored as a small graph file plus an &lt;strong&gt;external data&lt;/strong&gt; file holding the tensor payload. The graph does not know whether the bytes it points at are the ones you exported. Get the offsets wrong, or ship a re-exported payload, and &lt;code&gt;onnx.load()&lt;/code&gt; still succeeds, the session still runs, and the model still returns an image. It just returns &lt;em&gt;slightly different&lt;/em&gt; numbers, which is the hardest kind of regression to notice by eye.&lt;/p&gt;

&lt;p&gt;So the check has to be on the payload, not on the session.&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;import&lt;/span&gt; &lt;span class="n"&gt;hashlib&lt;/span&gt;
&lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="n"&gt;onnx&lt;/span&gt;
&lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="n"&gt;onnx&lt;/span&gt; &lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="n"&gt;numpy_helper&lt;/span&gt;

&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;tensor_digest&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;path&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;str&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;-&amp;gt;&lt;/span&gt; &lt;span class="kt"&gt;str&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;model&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;onnx&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;load&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;path&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;  &lt;span class="c1"&gt;# resolves external data into memory&lt;/span&gt;
    &lt;span class="n"&gt;h&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;hashlib&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;sha256&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;init&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;graph&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;initializer&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="n"&gt;arr&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;numpy_helper&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;to_array&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;init&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
        &lt;span class="n"&gt;h&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;update&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;init&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="nf"&gt;encode&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;span class="n"&gt;h&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;update&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;arr&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;tobytes&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;span class="k"&gt;return&lt;/span&gt; &lt;span class="n"&gt;h&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;hexdigest&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;Hashing the initializer names alongside the raw bytes stops a reordering of tensors from producing the same digest. For our size (about 178 MB of tensors) loading everything into memory is fine; for a multi-gigabyte model you would stream the external data file instead and hash the chunks.&lt;/p&gt;

&lt;h2&gt;
  
  
  The numbers we held ourselves to
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;178,397,184 bytes&lt;/strong&gt; of exact float32 tensor data, shared across the three Lite-2K graphs.&lt;/li&gt;
&lt;li&gt;Every complete ONNX ModelProto reconstructed byte-identically to its pinned export, after normalizing default storage metadata.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No quantization&lt;/strong&gt;, no graph geometry change, no extra segmentation step.&lt;/li&gt;
&lt;li&gt;Per-profile compressed distribution landed at 267 MB and 268 MB, against 760 MB for the previous combined runtime — roughly a 65% reduction in what gets pulled.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Those are the claims I am comfortable repeating because they are properties of files, not of a screenshot.&lt;/p&gt;

&lt;h2&gt;
  
  
  Byte-identical behaviour, not just byte-identical weights
&lt;/h2&gt;

&lt;p&gt;A matching hash of the weights does not prove the runtime assembles the graph the same way. We ran four real cases — the Lite profile, and the 2K profile on a portrait, a square and a landscape input — serially on Linux amd64 with 4 CPU and 7 GiB of memory, no swap. All four alpha arrays matched byte-for-byte, and so did both the JPEG and PNG encodings.&lt;/p&gt;

&lt;p&gt;Peak resident memory was 1.928 GiB for Lite and 5.790 GiB for Lite-2K. That gap is the entire argument for separate images: the 2K profile is what sets the memory budget, and nobody processing a small photo should pay it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Warming happens per container, not per profile
&lt;/h2&gt;

&lt;p&gt;A container that has just started holds no model. In a real 2K Flask startup, the first square warm loaded only the 2048-square graph, in 1.81 seconds. A repeat warm took 0.001 seconds and kept the same session. The landscape graph did not load at all.&lt;/p&gt;

&lt;p&gt;That is a one-sample observation, not a benchmark, but it shaped the code: a batch warms &lt;strong&gt;only its first input per profile&lt;/strong&gt;, because each container retains one session and the graph is chosen from the input's dimensions. Later orientation changes inside the same batch can still load a second graph — so "warmed" is a per-container fact that expires.&lt;/p&gt;

&lt;h2&gt;
  
  
  A bug the verification exposed
&lt;/h2&gt;

&lt;p&gt;One of the regression images was a rotated JPEG, and it failed. The parser was matching an uppercase EXIF signature and recognised nothing else, so standard Exif bytes were ignored and the image went in with the wrong orientation hint. The browser and server were already rotating pixels correctly; the parser disagreed with them. That is exactly the class of bug that a "the model is unchanged" check alone would have shipped.&lt;/p&gt;

&lt;h2&gt;
  
  
  What all of this still does not prove
&lt;/h2&gt;

&lt;p&gt;Local CPU samples are not Cloudflare performance. We verified production separately, with two anonymous tasks: a 768×1024 image on Lite warmed in 902 ms and completed in 16.7 s; a 1440×1440 image on Lite-2K warmed in 1454 ms and completed in 29.9 s. Those are single, network-inclusive samples, not P95s, and I would not quote them as latency guarantees.&lt;/p&gt;

&lt;p&gt;And byte-identical output is not an accuracy study. It says the new packaging behaves like the old one. It says nothing about whether the old one was good.&lt;/p&gt;

&lt;h2&gt;
  
  
  The checklist I now run for any model or packaging change
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;Pin the source revision and hash the payload, not the filename.&lt;/li&gt;
&lt;li&gt;Reconstruct the model and compare tensor digests against the pinned export.&lt;/li&gt;
&lt;li&gt;Run real inputs across every geometry the runtime can select, and compare output arrays &lt;em&gt;and&lt;/em&gt; encodings.&lt;/li&gt;
&lt;li&gt;Measure peak RSS per profile, on the smallest machine you intend to support.&lt;/li&gt;
&lt;li&gt;Treat local timings as smoke tests. Measure the deployed thing separately.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Disclosure: I build &lt;a href="https://cleanwhitebg.com" rel="noopener noreferrer"&gt;CleanWhiteBG&lt;/a&gt;, the product these container images serve — a browser tool that turns product photos into white-background JPGs or transparent PNGs.&lt;/p&gt;

</description>
      <category>onnx</category>
      <category>machinelearning</category>
      <category>python</category>
    </item>
    <item>
      <title>cosskill update: persona pages now teach the rewrite, not just the chat</title>
      <dc:creator>neo xia</dc:creator>
      <pubDate>Wed, 04 Nov 2026 17:27:05 +0000</pubDate>
      <link>https://dev.to/neo_xia_3f4c019330af5fb9d/cosskill-update-persona-pages-now-teach-the-rewrite-not-just-the-chat-318g</link>
      <guid>https://dev.to/neo_xia_3f4c019330af5fb9d/cosskill-update-persona-pages-now-teach-the-rewrite-not-just-the-chat-318g</guid>
      <description>&lt;p&gt;Two weeks ago I posted about the front-door change: cosskill — an AI roleplay simulator for rehearsing hard conversations — now opens the homepage straight into the ex-partner persona chat, and the Chinese site mirrors the same default.&lt;/p&gt;

&lt;p&gt;This week's quiet update is about teaching, not chatting: persona pages now carry a guided practice example. Instead of only describing who a persona is and what they believe, each page shows one complete rehearsal loop:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Draft — the message most people would actually send ("You have become careless. I should not have to keep reminding you.")&lt;/li&gt;
&lt;li&gt;Challenge — what the persona pushes back on: what exactly was missing, what was the impact, and what got in the way?&lt;/li&gt;
&lt;li&gt;Revision — the rewrite that names a standard and an observable gap instead of a character judgment ("The last handover lacked rollback instructions, so the release team had to pause. What got in the way? Before the next release, can we agree who checks that the instructions are complete?")&lt;/li&gt;
&lt;li&gt;Takeaway — why the revision works, with the feedback framework it follows cited by name&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The example also states its own limits: it is an illustrative, AI-assisted exercise — not a recorded session, not a quote from anyone, not evidence of an outcome.&lt;/p&gt;

&lt;p&gt;The point: you can read a persona's philosophy, then watch a bad message become a good one, before you run the same loop on your own conversation.&lt;/p&gt;

&lt;p&gt;Everything else stands: 5 messages a day free without signup, 10 a day with a free account, Pro $12.90/month if you need more. The homepage try-it-now still drops you straight into the ex persona.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://cosskill.com" rel="noopener noreferrer"&gt;https://cosskill.com&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>productivity</category>
      <category>career</category>
    </item>
    <item>
      <title>cosskill update: the hardest conversation is now the front door</title>
      <dc:creator>neo xia</dc:creator>
      <pubDate>Wed, 28 Oct 2026 17:29:00 +0000</pubDate>
      <link>https://dev.to/neo_xia_3f4c019330af5fb9d/cosskill-update-the-hardest-conversation-is-now-the-front-door-4oe2</link>
      <guid>https://dev.to/neo_xia_3f4c019330af5fb9d/cosskill-update-the-hardest-conversation-is-now-the-front-door-4oe2</guid>
      <description>&lt;p&gt;Two weeks ago cosskill added screenshot attachments to practice conversations. This week we changed the front door: the homepage "Try It Now" button no longer opens a persona menu — it opens a conversation with the ex-partner persona by default.&lt;/p&gt;

&lt;p&gt;Why change the default? Picking a persona is a decision, and decisions stall. The new default removes the stall: the first click now starts the conversation many people have been rehearsing in their head for months, and switching to any of the other personas is still one click away.&lt;/p&gt;

&lt;p&gt;Everything else is unchanged:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Free tier&lt;/strong&gt;: 5 messages a day without an account, 10 a day after signing in — resets daily, no credit card.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Pro&lt;/strong&gt;: $12.90/month or $129/year (regular $154.80/yr, ~17% off) for unlimited practice.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Screenshots&lt;/strong&gt;: attach images (up to 5 MB) to keep the replies worth keeping — attachments are retained for 30 days.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Languages&lt;/strong&gt;: English and Chinese (中文) interfaces.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;cosskill is an AI conversation simulator for rehearsing the conversations people dread — breakups, salary negotiation, performance reviews, saying no. If there is a conversation you keep postponing, the front door now assumes it is that one.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://cosskill.com" rel="noopener noreferrer"&gt;https://cosskill.com&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>productivity</category>
      <category>career</category>
    </item>
    <item>
      <title>cosskill update: attach screenshots to your practice conversations</title>
      <dc:creator>neo xia</dc:creator>
      <pubDate>Thu, 22 Oct 2026 02:08:48 +0000</pubDate>
      <link>https://dev.to/neo_xia_3f4c019330af5fb9d/cosskill-update-attach-screenshots-to-your-practice-conversations-2522</link>
      <guid>https://dev.to/neo_xia_3f4c019330af5fb9d/cosskill-update-attach-screenshots-to-your-practice-conversations-2522</guid>
      <description>&lt;p&gt;Quick update for anyone using &lt;a href="https://cosskill.com/" rel="noopener noreferrer"&gt;cosskill&lt;/a&gt; to rehearse high-stakes conversations — you can now attach screenshots to a practice conversation.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What's new&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;An &lt;strong&gt;attach-image button&lt;/strong&gt; in the chat: drop in a screenshot (a real email, a performance-review draft, a job posting, a message you're stuck on) and the AI persona works from it.&lt;/li&gt;
&lt;li&gt;Images are sent to the AI model and kept for up to 30 days in that chat, so a rehearsal session can refer back to what you shared.&lt;/li&gt;
&lt;li&gt;Made for the moments where the context &lt;em&gt;is&lt;/em&gt; the hard part — negotiating against an actual offer letter reads very differently from describing it in words.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Everything else stays the same:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;5 messages a day without an account&lt;/strong&gt; — no signup, no credit card&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;10 messages a day after signing in&lt;/strong&gt; (free), resetting daily&lt;/li&gt;
&lt;li&gt;All personas and scenarios are available on the free tier — salary negotiations, critical feedback, difficult personal talks&lt;/li&gt;
&lt;li&gt;The Chinese-language interface is fully live&lt;/li&gt;
&lt;li&gt;Pro, for unlimited practice, is &lt;strong&gt;$12.90/month or $129/year&lt;/strong&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If a tense conversation is coming up, bring the real artifact and rehearse against it: &lt;a href="https://cosskill.com/" rel="noopener noreferrer"&gt;https://cosskill.com/&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>productivity</category>
      <category>career</category>
    </item>
    <item>
      <title>Batch-processing product photos to a clean white background: a 3-pass workflow</title>
      <dc:creator>neo xia</dc:creator>
      <pubDate>Mon, 19 Oct 2026 16:11:41 +0000</pubDate>
      <link>https://dev.to/neo_xia_3f4c019330af5fb9d/batch-processing-product-photos-to-a-clean-white-background-a-3-pass-workflow-2gc7</link>
      <guid>https://dev.to/neo_xia_3f4c019330af5fb9d/batch-processing-product-photos-to-a-clean-white-background-a-3-pass-workflow-2gc7</guid>
      <description>&lt;p&gt;If you shoot a whole catalog in one afternoon, editing one photo at a time is where the evening goes. Here is the batch workflow we use to take a folder of product shots to storefront-ready white backgrounds — three passes, most of it automated.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Pass 1 — triage the folder (2 minutes)&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Open the folder and sort by shooting setup, not by product. All photos from the same light and angle belong in one batch: a tool that keys a white background calibrates per batch, and mixed setups in one batch produce inconsistent edges. Delete obvious rejects now (motion blur, lens cap still on) instead of paying processing time on them.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Pass 2 — batch the background pass&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Run the batch through a background-removal pass in one go, with the same output settings for the whole folder: white (#FFFFFF) output, one resolution, one quality level. Consistency is the point — a catalog where every image has the same white and the same framing reads as a store, not a flea market.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Pass 3 — edge QA on the risky 10%&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Not every photo deserves equal attention. Jewelry, glassware, and anything with hair-thin strands produce the most edge halos — pull those out and inspect at 200% zoom. Small fringes clean up with a 1-pixel edge contraction on white; anything worse goes back for a reshoot.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;File hygiene that saves an afternoon later&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Name outputs by product identifier plus angle suffix (sku-01-front.jpg), keep them ASCII, and export at the resolution the marketplace actually displays. A consistent naming scheme is what makes the next batch a ten-minute job instead of a search party.&lt;/p&gt;

&lt;p&gt;Disclosure: I build CleanWhiteBG (&lt;a href="https://cleanwhitebg.com" rel="noopener noreferrer"&gt;https://cleanwhitebg.com&lt;/a&gt;), a free browser tool that removes image backgrounds and replaces them with white — batch uploads need no signup, and every upload auto-deletes within 2 hours. The workflow above is the same one we run our own screenshots through. Nothing here is sponsored.&lt;/p&gt;

</description>
      <category>tutorial</category>
      <category>ecommerce</category>
      <category>productivity</category>
    </item>
    <item>
      <title>cosskill update: the free tier now resets daily</title>
      <dc:creator>neo xia</dc:creator>
      <pubDate>Wed, 14 Oct 2026 17:30:27 +0000</pubDate>
      <link>https://dev.to/neo_xia_3f4c019330af5fb9d/cosskill-update-the-free-tier-now-resets-daily-2di8</link>
      <guid>https://dev.to/neo_xia_3f4c019330af5fb9d/cosskill-update-the-free-tier-now-resets-daily-2di8</guid>
      <description>&lt;p&gt;Quick update for anyone using &lt;a href="https://cosskill.com/" rel="noopener noreferrer"&gt;cosskill&lt;/a&gt; to rehearse the conversations they dread — a salary negotiation, critical feedback, a difficult personal talk. The free tier just changed in a way that makes daily practice a lot easier.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The free tier now resets every day.&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;5 messages a day without an account&lt;/strong&gt; — no signup, no credit card, just open the site and start rehearsing&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;10 messages a day after signing in&lt;/strong&gt; (free)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The counter resets on a daily cycle, so one heavy day of practice no longer eats the whole budget. Come back tomorrow and keep going.&lt;/p&gt;

&lt;p&gt;Everything else stays the same:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;All personas and scenarios are available on the free tier — salary negotiations, feedback conversations, and difficult personal talks&lt;/li&gt;
&lt;li&gt;The Chinese-language interface (launched at the end of September) is fully live&lt;/li&gt;
&lt;li&gt;Pro, if you want unlimited practice, is &lt;strong&gt;$12.90/month or $129/year&lt;/strong&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Why the change? Rehearsal works as a habit, not a one-off. Preparing for a real negotiation next week means running the conversation a few minutes every day — with a persona that pushes back, interrupts, and doesn't grade on a curve — not burning through a fixed allowance the night before. A daily reset matches how people actually prepare.&lt;/p&gt;

&lt;p&gt;If you have a high-stakes conversation coming up, try it here: &lt;a href="https://cosskill.com/" rel="noopener noreferrer"&gt;https://cosskill.com/&lt;/a&gt;&lt;/p&gt;

</description>
      <category>career</category>
      <category>ai</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Auto-deleting user uploads within 2 hours: how a privacy promise becomes infrastructure</title>
      <dc:creator>neo xia</dc:creator>
      <pubDate>Sun, 11 Oct 2026 16:06:31 +0000</pubDate>
      <link>https://dev.to/neo_xia_3f4c019330af5fb9d/auto-deleting-user-uploads-within-2-hours-how-a-privacy-promise-becomes-infrastructure-36p9</link>
      <guid>https://dev.to/neo_xia_3f4c019330af5fb9d/auto-deleting-user-uploads-within-2-hours-how-a-privacy-promise-becomes-infrastructure-36p9</guid>
      <description>&lt;p&gt;When a privacy policy says "every upload is deleted within two hours," that sentence has to be backed by infrastructure, not intentions. This post walks through what implementing that promise actually looked like for a browser-based image tool — and the trade-offs nobody mentions when the roadmap says "privacy-first."&lt;/p&gt;

&lt;h2&gt;
  
  
  Why a hard retention window matters
&lt;/h2&gt;

&lt;p&gt;Image tools touch sensitive casual uploads: product shots before launch, personal photos, screenshots with private data. A retention promise is only credible when deletion is enforced by the storage layer, not by a cleanup job someone can forget to deploy.&lt;/p&gt;

&lt;h2&gt;
  
  
  Make the storage layer enforce the deadline
&lt;/h2&gt;

&lt;p&gt;The rule we landed on: nothing outlives its usefulness window by default.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Ephemeral object storage.&lt;/strong&gt; Every upload lands in object storage with a lifecycle rule that expires the object two hours after creation. Deletion is not scheduled application code — it is a property of the storage class. If the application dies, the objects still expire.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Process in memory wherever possible.&lt;/strong&gt; The background-removal pipeline runs per request; intermediate frames never touch persistent disk.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No analytics on image content.&lt;/strong&gt; Logs carry event types and timings, never pixels or filenames.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  The trade-offs
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;No "your recent uploads" gallery.&lt;/strong&gt; A server-side history would break the model. A session-scoped history held client-side is the honest version.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Support gets harder.&lt;/strong&gt; When a user reports a broken output, the artifact may already be gone. You learn to log structured errors instead of keeping evidence.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Lifecycle rules are eventually consistent.&lt;/strong&gt; "Within 2 hours" is an upper bound, not a precise instant — say exactly that in the policy, in the same units everywhere (no "60 minutes" vs "2 hours" drift).&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  The quarterly re-audit checklist
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Do lifecycle rules cover every bucket, including any replica?&lt;/li&gt;
&lt;li&gt;Do logs contain metadata only?&lt;/li&gt;
&lt;li&gt;Can the deletion promise be verified from the outside (policy page vs. observed behavior)?&lt;/li&gt;
&lt;li&gt;Is the retention window stated in identical units across product, policy, and docs?&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Disclosure
&lt;/h2&gt;

&lt;p&gt;Disclosure: I build CleanWhiteBG (&lt;a href="https://cleanwhitebg.com" rel="noopener noreferrer"&gt;https://cleanwhitebg.com&lt;/a&gt;), a free browser tool that removes image backgrounds and replaces them with white. This post comes from engineering that product; nothing here is sponsored. The user-facing summary of the discipline above: upload without signup, get a white-background result, and every upload auto-deletes within 2 hours.&lt;/p&gt;

</description>
      <category>architecture</category>
      <category>backend</category>
      <category>infrastructure</category>
      <category>privacy</category>
    </item>
    <item>
      <title>cosskill update: the AI conversation simulator now speaks Chinese</title>
      <dc:creator>neo xia</dc:creator>
      <pubDate>Wed, 07 Oct 2026 17:44:49 +0000</pubDate>
      <link>https://dev.to/neo_xia_3f4c019330af5fb9d/cosskill-update-the-ai-conversation-simulator-now-speaks-chinese-3fn6</link>
      <guid>https://dev.to/neo_xia_3f4c019330af5fb9d/cosskill-update-the-ai-conversation-simulator-now-speaks-chinese-3fn6</guid>
      <description>&lt;p&gt;cosskill is an AI communication simulator for rehearsing high-stakes conversations — salary negotiations, performance feedback, difficult personal talks — by speaking with realistic AI personas before they happen.&lt;/p&gt;

&lt;p&gt;This month we shipped a full Chinese-language interface. Every persona, scenario and settings screen now works in Chinese as well as English, so you can rehearse in the language you'll actually be speaking when the conversation matters.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What you can do without an account:&lt;/strong&gt; 3 full messages, every persona and scenario unlocked. &lt;strong&gt;After signing in:&lt;/strong&gt; 10 messages. Pro removes the cap entirely — $12.90/month, or $9.90/month billed yearly.&lt;/p&gt;

&lt;p&gt;Why rehearse at all? Negotiations are usually decided in the first two minutes: the opening ask, the framing, the pause after a lowball counter. Practicing out loud against an AI that pushes back realistically is the cheapest way to find the weak spots in your opening — before a real salary is on the table.&lt;/p&gt;

&lt;p&gt;Try it: &lt;a href="https://cosskill.com" rel="noopener noreferrer"&gt;https://cosskill.com&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>productivity</category>
      <category>career</category>
    </item>
    <item>
      <title>Why marketplace product photos get rejected: white-background image requirements, and a 1-minute browser fix</title>
      <dc:creator>neo xia</dc:creator>
      <pubDate>Sun, 04 Oct 2026 16:00:57 +0000</pubDate>
      <link>https://dev.to/neo_xia_3f4c019330af5fb9d/why-marketplace-product-photos-get-rejected-white-background-image-requirements-and-a-1-minute-38md</link>
      <guid>https://dev.to/neo_xia_3f4c019330af5fb9d/why-marketplace-product-photos-get-rejected-white-background-image-requirements-and-a-1-minute-38md</guid>
      <description>&lt;p&gt;If you sell on marketplaces — Amazon, Shopify, Etsy, app stores, or any product directory — you have probably had a listing image rejected or down-ranked for one reason: the background.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Disclosure:&lt;/strong&gt; I build &lt;a href="https://cleanwhitebg.com" rel="noopener noreferrer"&gt;CleanWhiteBG&lt;/a&gt;, a free browser tool for exactly this problem. This post explains the requirements themselves; the tool is one way to hit them. No signup, no watermark, and images never leave your device.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why "white background" is a hard requirement
&lt;/h2&gt;

&lt;p&gt;Most marketplaces normalize their catalog grids around a consistent look. A white (or pure transparent) background lets them:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Composite your photo into any layout&lt;/strong&gt; — tiles, carousels, comparison tables — without visible seams.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Pass automated checks&lt;/strong&gt;: several pipelines measure background uniformity and reject images with busy or colored backgrounds.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Keep perceived quality high&lt;/strong&gt;: studio-white reads as "professional" at thumbnail size, where most buying decisions happen.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What the requirements usually say
&lt;/h2&gt;

&lt;p&gt;The wording varies, but the checks look like this:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;The product fills 80–90% of the frame.&lt;/li&gt;
&lt;li&gt;The background is pure white (#FFFFFF) or transparent PNG.&lt;/li&gt;
&lt;li&gt;No watermarks, logos, or inset text on the image.&lt;/li&gt;
&lt;li&gt;Sharp edges — no halo or color fringing left behind by a bad cutout.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Point 4 is where DIY jobs fail: a tolerance-based "magic wand" removal leaves gray fringes around hair, glassware, or jewelry, and reviewers (human or automated) notice.&lt;/p&gt;

&lt;h2&gt;
  
  
  The 1-minute fix in the browser
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;Open &lt;a href="https://cleanwhitebg.com" rel="noopener noreferrer"&gt;https://cleanwhitebg.com&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Drop the product photo in (nothing is uploaded — it runs locally in your browser via WebGPU/WASM).&lt;/li&gt;
&lt;li&gt;Pick the mode: &lt;strong&gt;Fast&lt;/strong&gt; for everyday shots, &lt;strong&gt;Precision&lt;/strong&gt; for complex edges like hair or jewelry.&lt;/li&gt;
&lt;li&gt;Download the flattened white-background PNG and upload it to the marketplace.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Because the processing is local, it also works for images you would rather not upload to a cloud service — product shots under NDA, unreleased packaging, client work.&lt;/p&gt;

&lt;h2&gt;
  
  
  Checklist before you re-upload
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;[ ] Product fills the frame, centered&lt;/li&gt;
&lt;li&gt;[ ] Background is exactly #FFFFFF (pick the white-flatten option, not transparent)&lt;/li&gt;
&lt;li&gt;[ ] Zoom to 200% and check edges for fringing&lt;/li&gt;
&lt;li&gt;[ ] No leftover watermark from stock previews&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;That is the whole loop. If you batch 50 SKUs a week, the difference between a 5-minute-per-image desktop workflow and a 30-second browser one adds up fast.&lt;/p&gt;

&lt;p&gt;Feedback welcome — especially edge cases where the cutout struggles (glass, fur, reflective packaging).&lt;/p&gt;

</description>
      <category>design</category>
      <category>productivity</category>
      <category>tools</category>
    </item>
    <item>
      <title>Four Marketplaces, One Listing Workflow: What Amazon, eBay, Etsy and Shopify Expect From Your Copy</title>
      <dc:creator>neo xia</dc:creator>
      <pubDate>Sat, 03 Oct 2026 19:38:42 +0000</pubDate>
      <link>https://dev.to/neo_xia_3f4c019330af5fb9d/four-marketplaces-one-listing-workflow-what-amazon-ebay-etsy-and-shopify-expect-from-your-copy-3n8c</link>
      <guid>https://dev.to/neo_xia_3f4c019330af5fb9d/four-marketplaces-one-listing-workflow-what-amazon-ebay-etsy-and-shopify-expect-from-your-copy-3n8c</guid>
      <description>&lt;p&gt;If you sell on more than one marketplace, you already know the copy you write for Amazon doesn't survive a trip to Etsy. Same product, different rules — and the rules are not cosmetic.&lt;/p&gt;

&lt;p&gt;I build listing tools for a living, and here's the cheat sheet I wish someone had handed me before I rewrote the same SKU four times.&lt;/p&gt;

&lt;h2&gt;
  
  
  What each platform actually expects
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Amazon&lt;/strong&gt; — Titles get truncated hard on mobile, and Rufus surfaces content that answers buyer questions directly. Front-load the searchable keywords; save the adjectives.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;eBay&lt;/strong&gt; — Title length is capped at 80 characters, and buyers filter by item specifics. Condition-forward descriptions reduce disputes.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Etsy&lt;/strong&gt; — Buyers respond to the maker story and materials. Tags are multi-word phrases, and the tone is handmade, not corporate.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Shopify&lt;/strong&gt; — Your own brand voice applies, and the most common failure is shipping the manufacturer's thin spec sheet as a product description.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why paste-reformatting breaks down
&lt;/h2&gt;

&lt;p&gt;The usual loop looks like this: ask a general AI writer for copy, paste the answer into the marketplace form, then fix the character limit, reorder the bullets, rewrite the description, and swap images by hand. Every SKU repeats the loop, and the loop scales with your catalog.&lt;/p&gt;

&lt;h2&gt;
  
  
  What we built instead
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://aiproductpageoptimization.com" rel="noopener noreferrer"&gt;AI Product Page Optimization&lt;/a&gt; is a Chrome extension that reads the live listing page you already have open, detects which marketplace you're on, and generates copy that respects that platform's rules. One install covers Amazon, Shopify (including custom domains), eBay, and Etsy. Optional image enhancement runs in the same pass, and results land on a review dashboard so nothing publishes until you approve it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Pricing facts
&lt;/h2&gt;

&lt;p&gt;Text optimization costs 1 credit; successfully enhanced images cost 2. New accounts get 5 free credits, no card required. Subscriptions start at $19.9/month.&lt;/p&gt;

&lt;p&gt;If you manage listings across channels: which platform's rules eat most of your time? Comments are open.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>ecommerce</category>
      <category>shopify</category>
      <category>marketing</category>
    </item>
    <item>
      <title>Ship an Editor, Not a Generator: Designing Review-First AI for Marketplace Listings</title>
      <dc:creator>neo xia</dc:creator>
      <pubDate>Fri, 02 Oct 2026 19:25:32 +0000</pubDate>
      <link>https://dev.to/neo_xia_3f4c019330af5fb9d/ship-an-editor-not-a-generator-designing-review-first-ai-for-marketplace-listings-151d</link>
      <guid>https://dev.to/neo_xia_3f4c019330af5fb9d/ship-an-editor-not-a-generator-designing-review-first-ai-for-marketplace-listings-151d</guid>
      <description>&lt;p&gt;Most AI writing tools optimize for one moment: the instant text appears. But for marketplace sellers, that moment is the least important part of the workflow. What matters is everything after — reading the suggestion, comparing it to the original, and deciding what actually goes live on Amazon, Shopify, eBay, or Etsy.&lt;/p&gt;

&lt;p&gt;That gap is why we built &lt;a href="https://aiproductpageoptimization.com" rel="noopener noreferrer"&gt;AI Product Page Optimization&lt;/a&gt; around an editor, not a generator.&lt;/p&gt;

&lt;h2&gt;
  
  
  The trust problem with generated text
&lt;/h2&gt;

&lt;p&gt;A generated title is a guess. A reviewed title is a decision. Sellers don't publish what an AI writes — they publish what they approve. So the core UX question isn't "how good is the model?" but "how fast can a seller judge the output?"&lt;/p&gt;

&lt;p&gt;Our answer: every optimization run lands on a web dashboard as a before/after comparison. Original on one side, suggestion on the other. The seller edits, accepts, or discards. Nothing touches the listing until a human says so.&lt;/p&gt;

&lt;h2&gt;
  
  
  The loop: extract → draft → review → publish
&lt;/h2&gt;

&lt;p&gt;The Chrome extension reads the live product page the seller already has open — Amazon, Shopify (including custom domains), eBay, or Etsy — using structured data plus platform-specific selectors. The AI drafts rewrites that respect each channel's constraints: Amazon's mobile truncation, eBay title length, Etsy's handmade tone, Shopify brand voice.&lt;/p&gt;

&lt;p&gt;Then control moves to the dashboard:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Before/after comparison for every field&lt;/li&gt;
&lt;li&gt;History of past runs, reusable across SKUs&lt;/li&gt;
&lt;li&gt;Optional public share links for team feedback&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Paste-ready output, human-approved.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why the same pass handles images
&lt;/h2&gt;

&lt;p&gt;Copy alone rarely fixes a weak listing. The same run can enhance product photos or swap backgrounds — and keeping the original image is free. That pricing detail matters more than it sounds: it makes "do nothing" a legitimate choice, which keeps the review loop honest.&lt;/p&gt;

&lt;h2&gt;
  
  
  What we learned
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Constraints are features.&lt;/strong&gt; Character limits and platform tone rules make suggestions &lt;em&gt;more&lt;/em&gt; useful, not less — they turn generated text into paste-ready text.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Review speed beats generation speed.&lt;/strong&gt; A suggestion that takes 30 seconds to judge is worth more than three alternatives that take 10 minutes to compare.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;History turns AI output into an asset.&lt;/strong&gt; Sellers who can reopen last month's runs stop treating AI as a slot machine.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;If you sell across marketplaces and want to see the review-first flow yourself, the extension is free to try — 5 credits on signup, no credit card: &lt;a href="https://aiproductpageoptimization.com" rel="noopener noreferrer"&gt;https://aiproductpageoptimization.com&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>webdev</category>
      <category>ux</category>
      <category>productivity</category>
    </item>
  </channel>
</rss>
