<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[Jack's Blog]]></title><description><![CDATA[Jack's Blog]]></description><link>https://seedancejack.hashnode.dev</link><generator>RSS for Node</generator><lastBuildDate>Mon, 16 Nov 2026 22:29:54 GMT</lastBuildDate><atom:link href="https://seedancejack.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[A Motion Note for a Pull Request from the Frames You Captured in Review]]></title><description><![CDATA[A visual change in a pull request is still, too often, a set of screenshots. The next reviewer has to reconstruct the path: which screen, which state…]]></description><link>https://seedancejack.hashnode.dev/a-motion-note-for-a-pull-request-from-the-frames-you-captured-in-review</link><guid isPermaLink="true">https://seedancejack.hashnode.dev/a-motion-note-for-a-pull-request-from-the-frames-you-captured-in-review</guid><category><![CDATA[pull request]]></category><category><![CDATA[code review]]></category><category><![CDATA[frontend]]></category><category><![CDATA[ai video]]></category><dc:creator><![CDATA[Jack]]></dc:creator><pubDate>Mon, 16 Nov 2026 06:51:09 GMT</pubDate><content:encoded><![CDATA[<p>A visual change in a pull request is still, too often, a set of screenshots. The next reviewer has to reconstruct the path: which screen, which state, which transition actually changed. I have started attaching a 15- to 20-second motion note to reviews that are hard to judge from stills — especially layout shifts, hover and empty states, and anything that only makes sense as a sequence. I do not record my whole screen and talk over it for three minutes. I build the note from the frames I already captured while I was looking at the diff, plus one cursor-path clip if I have it, and a single written caption.</p>
<p>The point is not a polished demo. The point is that a reviewer can see the sequence without checking out the branch or guessing which of five screenshots is the moment that matters.</p>
<h2>Collect the frames while you are still in the review</h2>
<p>I do not go back later and restage the UI for a recording. I keep a small set from the pass I already made:</p>
<ul>
<li>The entry state: the screen as it looks before the new interaction, at the default size, with the real copy.</li>
<li>The changed state: the same screen after the click, the hover, the resize, or the empty-state path the ticket describes.</li>
<li>One still of a related surface only if the change crosses a page — a modal, a drawer, a toast — and only if that surface is in the branch.</li>
<li>The before still, if the review is about a regression or a comparison, clearly marked as the previous behavior so it cannot be confused with the new one.</li>
<li>If I already recorded a few seconds of the cursor moving through the change, I keep that clip as a motion reference for pace and for the path, not as a five-minute screen capture.</li>
</ul>
<p>I throw out anything that should not leave my machine: production data, a customer's name, an access token in a URL, a staging banner, a console stack trace I do not need in the note. If a frame is useful except for one field, I redact it before I attach it.</p>
<p>I also write down the facts the clip is not allowed to change: the component name, the copy in the UI, whether the state is empty or filled, and the breakpoint I used. A motion note that invents a button or smooths away a layout bug is worse than the original screenshots. This is a review artifact, not a launch film.</p>
<h2>Describe the sequence the reviewer has to see</h2>
<p>I inventory the files at the top of the brief, in the order I will attach them: these stills are the states, in the order the user hits them; this clip, if I have it, is only the pace of the cursor through the change. Then I describe a single path of 15 to 20 seconds. A typical version looks like this:</p>
<ul>
<li>Hold on the entry state long enough to read the screen.</li>
<li>Move through the interaction at the pace of a real click, not a jump cut.</li>
<li>Hold on the changed state, then on the related surface if there is one.</li>
<li>End on a quiet beat at the final state, with no extra chrome and no change to the copy.</li>
</ul>
<p>I stay inside the limits of the tool I use, <a href="https://tryseedance3.com/">Seedance 3.0</a>: one continuous take of up to 30 seconds, at up to 1080p, with the option to attach up to 50 image, video, and audio references in the same job. Four to eight stills and one motion clip are plenty. If I cannot fit the whole path in 20 seconds without rushing, I cut the path. A calm move from the entry state, through the click, to the result is more useful than a tour of every related screen in the app.</p>
<p>Sound stays off. I ask for a quiet interface, no music, no invented narrator, and no click track that was not recorded. The caption on the pull request carries the ticket number, the breakpoint, and the one sentence that states what changed. Generated speech in a review note is a liability: it can misstate the behavior, and it is harder to skim than a line of text.</p>
<h2>The review I do before I attach the note</h2>
<p>I watch the take twice. First I look only for product facts: did a label drift, did a state appear that is not in the branch, did a layout bug get smoothed over? If any of that happens, the take is not a candidate, no matter how smooth the move is. Then I look at the path: does it follow the ticket, is the pace slow enough to judge the transition, does it hold on the changed state?</p>
<p>When the take is almost right, I do not start a new brief from scratch. I attach the almost-right clip and describe only the fix: keep this path and this pace; restore the empty-state copy to the screenshot; do not change the spacing. That habit is what keeps a motion note from turning into a second review pass.</p>
<p>I export one file, small enough to attach or to link from the pull request, and I name it after the ticket, not after the tool I used. If a take looks cleaner than the screenshots but disagrees with them, I keep the screenshots and throw the take away. The branch is the source of truth; the clip is only a way to walk it.</p>
<p>The test I use is simple. A reviewer who has not pulled the branch should be able to say what changed, in what order, and what did not change. If that fails, the stills can carry the review until I have another pass.</p>
]]></content:encoded></item><item><title><![CDATA[Keep a Docs Mascot Consistent: Character Locks and Multi-Image Combines]]></title><description><![CDATA[A documentation site can survive a slightly ugly screenshot. It cannot survive a mascot that changes species between chapters. We learned this the slow…]]></description><link>https://seedancejack.hashnode.dev/keep-a-docs-mascot-consistent-character-locks-and-multi-image-combines</link><guid isPermaLink="true">https://seedancejack.hashnode.dev/keep-a-docs-mascot-consistent-character-locks-and-multi-image-combines</guid><category><![CDATA[frontend]]></category><category><![CDATA[Design]]></category><category><![CDATA[documentation]]></category><category><![CDATA[workflow]]></category><dc:creator><![CDATA[Jack]]></dc:creator><pubDate>Sat, 14 Nov 2026 08:37:40 GMT</pubDate><content:encoded><![CDATA[<p>A documentation site can survive a slightly ugly screenshot. It cannot survive a mascot that changes species between chapters. We learned this the slow way. We had a small illustrated character — a round bird with a tool belt — that was supposed to appear in getting-started pages, error-state illustrations, and a couple of social cards. After three months we had four birds. The beak got sharper. The belt became a satchel. One chapter's bird had eyebrows the others did not.</p>
<p>The writing team was not sloppy. We were describing the bird in prose and hoping each new scene would land on the same animal. That is not how these models work, and it is not how I would version a logo. The fix was to treat the mascot like source, and to build every new scene as a combine of that source plus a setting, inside an <a href="https://ai-image-editor.com/">AI image editor</a> that can hold character consistency and blend more than one reference in a single job.</p>
<h2>Put the character in a repository, not in a paragraph</h2>
<p>The file that matters is not the last PNG we published. It is the reference sheet.</p>
<pre><code>docs-art/
  refs/
    mascot-front.png
    mascot-three-quarter.png
    mascot-expression-ok.png
    mascot-expression-error.png
    set-desk.png
    set-terminal-error.png
    ui-empty-state.png
  prompts/
    getting-started-desk.txt
    error-state.txt
    og-card.txt
  out/
</code></pre>
<p><code>mascot-front.png</code> is the lock for the face and body. The three-quarter and expression sheets exist because a getting-started bird looking at a laptop is not the same pose as an error-state bird looking at a red badge, and I do not want the model to invent a fifth expression. The set files are places, not decorations: a desk we already photograph for the product, a terminal crop, an empty-state UI.</p>
<p>In the prompt I tag each file by role, the same way I would name layers.</p>
<blockquote>
<p>Use the first image as the character's face and body. Keep the tool belt, the eye shape, and the colors exactly. Use the second image only for the desk and window light. Place the bird at the desk, looking at a laptop, mid shot, soft morning light. Do not change the character's proportions or add a second bird.</p>
</blockquote>
<p>If I leave the roles unsaid, the desk texture ends up on the belt, or the bird inherits a plant from the set photo.</p>
<h2>Character consistency is a constraint, not a style</h2>
<p>The editor's character-consistency feature is the reason this works at all. The output is supposed to look like the input character, not like a cousin. I still write the lock in the prompt, because a constraint you do not state will be treated as optional the first time you ask for a dramatic scene.</p>
<p>Two rules have saved us from another flock of almost-birds:</p>
<p><strong>Do not update the character in the same job that changes the scene.</strong> If we decide the belt should have one more pocket, that is a revision of <code>mascot-front.png</code>, reviewed like a logo change. Then every scene prompt points at the new sheet. Mixing "new belt" and "now the bird is in a server room" is how you get a new bird in a new room and cannot tell which instruction did it.</p>
<p><strong>Do not use a previous scene as the only character lock.</strong> A scene already contains a desk, a pose, and lighting. If I feed that back as the identity reference, the next chapter inherits the desk. Identity references should be as clean as I can make them.</p>
<h2>Multi-image combination is the actual production step</h2>
<p>Most docs illustrations are not "a bird." They are "a bird plus this UI plus this room." The editor's multi-image combination is built for that blend. Depending on the model, you can attach up to fourteen references with Nano Banana 2.1, Nano Banana 2, and Nano Banana Pro, or up to sixteen with GPT Image 2.6 and GPT Image 2.5. I have never needed sixteen. I have often needed three: character, set, UI.</p>
<p>The UI file gets its own sentence, and it is a strict one.</p>
<blockquote>
<p>The laptop screen must show the empty-state UI from the third image, including the typography. Do not rewrite labels or invent extra buttons.</p>
</blockquote>
<p>Illustrated mascots and real product screens in one frame is a messy request. When the text on the screen starts to drift, I split the job: generate the bird-in-the-room with a blank or lightly blurred screen, then composite the real screenshot later, or run a second pass that only asks to place the UI. A single generate that has to invent a pose and perfectly typeset a settings page is usually doing too much.</p>
<p>For a wide docs banner I change only the framing. Most models support Auto and ratios from 1:1 to 21:9; Nano Banana 2.1 adds 4:1 and 8:1, which is where a thin banner actually fits. I do not crop a square mascot shot into an 8:1 strip and hope the beak stays on screen.</p>
<h2>What we accept, and what we send back</h2>
<p>The review looks more like a pull request than an art critique.</p>
<ul>
<li>The beak, belt, and colors match the reference sheet at a glance.</li>
<li>There is one bird. A helpful second bird in the corner is a failure.</li>
<li>The UI, if any, is still our UI.</li>
<li>The set does not introduce a logo or a product we do not ship.</li>
<li>The file is named for the page it belongs to, not "final-v7-really.png."</li>
</ul>
<p>When a take is almost right — good bird, wrong time of day — I describe only that change and reattach the same character sheet. Regenerating the whole scene from a looser prompt is how the eyebrows come back.</p>
<p>Outputs land in My Creations, where I download the ones we keep. I treat that library as a cache, not as the source of truth; the source of truth is <code>refs/</code> in the repository. If the public-visibility default worries you for unpublished brand work, check it before you start a batch, and prefer a setup where you can turn visibility off.</p>
<p>The measure of this workflow is not whether the bird looks expensive. It is whether chapter four's bird would be recognized by someone who only saw chapter one. Once that is true, adding a new page is a new set photo and a new prompt, not a new character design.</p>
]]></content:encoded></item><item><title><![CDATA[Treat AI Video Like a Build Artifact: References, Variants, and Platform Exports]]></title><description><![CDATA[Teams that publish a lot of product content already know how to treat copy and screenshots as source. The markdown is in git. The screenshots have names…]]></description><link>https://seedancejack.hashnode.dev/treat-ai-video-like-a-build-artifact-references-variants-and-platform-exports</link><guid isPermaLink="true">https://seedancejack.hashnode.dev/treat-ai-video-like-a-build-artifact-references-variants-and-platform-exports</guid><category><![CDATA[workflow]]></category><category><![CDATA[Productivity]]></category><category><![CDATA[ai video]]></category><category><![CDATA[content pipeline]]></category><dc:creator><![CDATA[Jack]]></dc:creator><pubDate>Fri, 13 Nov 2026 09:10:43 GMT</pubDate><content:encoded><![CDATA[<p>Teams that publish a lot of product content already know how to treat copy and screenshots as source. The markdown is in git. The screenshots have names. Video is the usual exception: a final file on a desktop, with every new ratio cropped and every new background reshot.</p>
<p>If you are using a generator that can lock references, emit a timed shot, and then edit that shot by description, you can treat video more like a build. This article is a layout and a set of rules that have survived a few months of shipping changelog videos, launch loops, and docs clips.</p>
<h2>Put identity in source, and renders in output</h2>
<p>I keep a content repository that looks like this:</p>
<pre><code>content/launch-webhooks-v2/
  brief.md
  refs/
    product-front.png
    product-label.png
    host.jpg
    set-studio.jpg
    grade-still.jpg
    hook-v2.wav
  prompts/
    master-16x9.txt
    vertical-9x16.txt
    night-diff.txt
  out/
    16x9-1080/
    9x16-1080/
    1x1-1080/
</code></pre>
<p><code>refs/</code> is source. It is reviewed like copy. If the product label changes, the file changes, and everything downstream is stale on purpose. <code>prompts/</code> is source too; the text that produced a take is as important as the take. <code>out/</code> is disposable. I never edit pixels in <code>out/</code> if I can describe the change instead, because a hand retouch cannot be replayed when the label updates.</p>
<p>The brief is short and boring on purpose: duration, aspect ratios required, the single action, the spoken line, and which refs are mandatory. If the brief needs a second action, it needs a second shot or a timestamped beat, not a longer adjective list.</p>
<h2>Make a master take, then derive variants as diffs</h2>
<p>The master is usually a 16:9 1080p shot that has the right bones: the product is readable, the camera path is decided, the line lands, there is a short hold. I do not try to win every platform in the first generate.</p>
<p>Variants I treat as diffs, not as new projects:</p>
<ul>
<li><strong>Framing variants.</strong> Same refs, same line, a prompt that changes only the aspect ratio and the safe area ("keep the bottle in the upper third for 9:16"). I would rather spend a render than crop a face off the edge.</li>
<li><strong>Look variants.</strong> Upload the master and ask for a change of weather, time of day, or art direction, keeping framing and motion. This is how a daytime docs clip becomes a darker launch teaser without a second shoot.</li>
<li><strong>SKU variants.</strong> Upload the master, attach the new packshot, and swap only the product. The table, the hands, and the camera stay.</li>
<li><strong>Length variants.</strong> If a paid placement needs a longer hold, extend the take forward rather than regenerating the hook.</li>
</ul>
<p>The rule is one kind of change per revision, the same way a pull request should not refactor and add a feature and change CI at once. When a variant fails, I want to know which instruction failed.</p>
<h2>Encode the platform matrix in the prompt, not in the editor</h2>
<p>A surprising amount of pain lives in last-mile crops. The generator I use can emit 16:9, 9:16, 4:3, 3:4, 1:1, 21:9, or follow the source still, at 480p, 720p, or 1080p. That means the matrix can be a list in <code>brief.md</code>:</p>
<table>
<thead>
<tr>
<th>Surface</th>
<th>Ratio</th>
<th>Resolution</th>
<th>Notes</th>
</tr>
</thead>
<tbody><tr>
<td>Docs / YouTube</td>
<td>16:9</td>
<td>1080p</td>
<td>Master; two-second hold</td>
</tr>
<tr>
<td>Shorts / Reels</td>
<td>9:16</td>
<td>1080p</td>
<td>Same line, tighter mid shot</td>
</tr>
<tr>
<td>In-app tile</td>
<td>1:1</td>
<td>720p</td>
<td>No dialogue, motion only</td>
</tr>
<tr>
<td>Blog hero</td>
<td>21:9</td>
<td>1080p</td>
<td>Locked product, no host</td>
</tr>
</tbody></table>
<p>Each row is a prompt file that inherits the same refs. The in-app tile, for example, is not a cropped master with the words cut off mid-syllable; it is a deliberate silent move around the same product lock.</p>
<h2>Name the jobs of up to dozens of files, or do not attach them</h2>
<p>It is tempting to empty a brand folder into a generation and hope consistency falls out. It does not. A useful upper bound — dozens of images, plus a smaller set of clips and audio files, in a single generation — only helps if every file has a job in the prompt: face, location, grade, product, motion guide, voice rhythm. I cap what I attach to what I can name. Unused refs are noise, and noise is how a grade still becomes a wardrobe.</p>
<p>Sound is part of the artifact, not a later phase, when the tool renders voices, room tone, and effects with the picture. If a surface cannot use dialogue, I say so in that prompt instead of stripping audio in five different editors later.</p>
<h2>The acceptance check I run like a test suite</h2>
<p>Before I merge a take into <code>out/</code> and hand it to marketing or docs:</p>
<ul>
<li>Pause on three frames: start, mid-move, hold. The product or face still matches the ref.</li>
<li>The prompt's single action is the action on screen.</li>
<li>The line, if any, is complete and roughly on the mouth.</li>
<li>Variants that were supposed to change one thing have not changed three.</li>
<li>File names in <code>out/</code> include ratio, resolution, and the prompt version.</li>
</ul>
<p>I run this system on <a href="https://tryseedance3.com/">Seedance 3.0</a> because it fits the artifact model: text plus tagged image, video, and audio references in one generation; a single shot of up to thirty seconds; sound in the same pass; and an edit path that restyles, swaps an element, or extends a take without throwing the master away. Commercial use is allowed on a paid plan, which matters once the output goes into a changelog or an ad.</p>
<p>The point is not to pretend a generator is a video team. It is to stop treating every new cut as a unique snowflake. If the refs and the prompt are source, and the render is a build, last week's launch video is a recipe you can run again when the label changes.</p>
]]></content:encoded></item></channel></rss>