Communitygithub.com

changelog-clips

Make one short UI demo per shipped change from a web app's real screens. Delivers separate silent web, square and portrait MP4/WebM clips, result posters and editable source. Use for changelog media, release-note loops and individual feature-update clips; use a whole release-film skill for a montage.

What is changelog-clips?

changelog-clips is a Cursor agent skill that make one short UI demo per shipped change from a web app's real screens. Delivers separate silent web, square and portrait MP4/WebM clips, result posters and editable source. Use for changelog media, release-note loops and individual feature-update clips; use a whole release-film skill for a montage.

Works with~Claude Code~Codex CLI✓Cursor
npx skills add https://github.com/omnigami-ai/plugins/tree/main/plugins/omnigami-launch-films/skills/changelog-clips

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

Changelog clips

Explicit user instructions take priority over this workflow. Work only on authorized products and sample data. Do not request passwords, API keys, payment data, government identifiers, health records or other sensitive data. Explain missing local dependencies before installing them in the working project. Network access is for the authorized product and its licensed assets, plus the declared npm dependency; respect access controls. Render one job at a time and wait for completion. Keep publishing, sending and deployment outside this task.

One native panel, one action, its held result, then an announced Demo reset. Keep the camera fixed and leave even padding around the whole panel. Read style.md before building and critique.md before judging.

Inputs and setup

Read the release notes and inspect the working product first. Collect only what they cannot supply: the changes to announce, a demo URL or local app, safe sample data, brand/font files and requested sizes. Version one covers visible web-app changes. Hold API-only changes rather than inventing a dashboard. Do not announce a feature as shipped on the strength of an upcoming preview.

Check Node 20+, ffmpeg and Chrome (scripts/chrome.mjs, or CHROME_PATH). Install puppeteer-core@^24 in the working project (it runs on Node 20; newer major versions can require a newer Node). SKILL_DIR below is the folder containing this file.

mkdir -p changelog-clips
cp -R "$SKILL_DIR/reference" changelog-clips/film
cp -R "$SKILL_DIR/scripts" changelog-clips/scripts
cp -R "$SKILL_DIR/music" changelog-clips/music
cd changelog-clips
npm init -y
npm install -D 'puppeteer-core@^24'

Make a separate film folder for each change; the reference is the appearance control, and examples/compare-matching-stays/ shows a native list. Remove the copied captures after studying their format and replace them with this product's. Remove copied out/ if present. Keep every film self-contained. On later releases, reuse the brand and unchanged clips; recapture only changed actions.

Product and story

Write PRODUCT.md with the actual product, brand, data permissions and source of each release claim. Write STORY.md with one entry per clip: stable ID, release source/date, title, before state, reachable route, real control/action, retained result, reset method, capture bounds and intended playback policy. Each independent clip must make sense beside its own release text. Do not combine two changes into one timeline.

Make the title agree with the demonstrated completion: choosing a replay point is not executing a replay, and previewing a parsed schedule is not saving one. Say “Choose…” or “Preview…” when that is the actual result; do not borrow the release's broader claim for a narrower demonstration.

Before committing to a change, measure its complete native component and decisive result text at the actual viewing size. Record the CSS dimensions, text size and resulting 358 px size in the story. A request for two clips does not make two suitable changes exist. Hold an incompatible item and say why; never deliver it as a finished phone/social clip with a small-print caveat. Select another genuinely documented change if one fits. When several shapes are requested and the user didn't name the change, measure every documented candidate against every requested shape before choosing, and prefer one whose complete component fits all of them; deliver a partial family only when none does. If none fits, return the product limitation without inventing UI or lowering the floor.

A component may fit only some shapes. Declare each unsupported shape and its measured reason in clip.json's heldShapes (for example {"4:3":"Native result is 10.7 px at 358 px; floor 12"}). The painter, preflight and exporter omit that held shape, while checking and delivering compatible ones. The manifest marks this partial; report every missing requested shape prominently. A partial family is not a passed three-shape proof. If every shape is held, deliver the measurements/report, not a movie.

Show that plan for review unless the user asked for unattended work. Default to 1200 × 900 web, 1080 × 1080 square and 1080 × 1350 portrait. These are separate compositions, not crops of one movie. A short stationary release title sits outside the UI on social; the web clip uses adjacent release text.

Capture and adapt

  1. Find one whole native card or panel that contains both the action and its result. Check its native narrow layout before capture. At 358 px viewing width the changed label, value or result must be about 12 px or larger. If the complete component cannot meet that, choose a compact view the app actually has or report the limitation. Never redraw or restyle it for the film.
  2. Copy/adapt film/plans/capture.json and its capture-setup.mjs. Set its URL, viewport, selectors, sample setup and inspect evidence. Record piece and target in rects.before, and again after the result. Capture whole-surface before, hover, pressed, changed (still hovered) and result (pointer cleared). Use real mouse down/up so the app supplies its own press. Run node scripts/centring.mjs film/plans/capture.json --no-crawl before capture.
  3. Run node scripts/screens.mjs film/plans/capture.json film/captures. Use animate for the app's own CSS/Web Animations frames. Include enough bounds for the largest native extent; the capture tool measures the region before the action. For an instant action, a short explicit duration records evidence with found: []: switch the whole capture instantly. For JS-driven motion the capture tool cannot seek, supply a deterministic native recording; never replace it with a fade or a motion assembled from stills.
  4. Copy the product's licensed font into film/fonts/, set film/clip.json and run node scripts/build-clip.mjs film film/clip.json. The options are shown below. It derives whole-surface geometry, cursor target and native-frame bindings from captures/screens.json, samples the native grounds, writes config.js and runs every shape's native-size checks. A failure holds the item: stop before exporting it. Review every capture before using it. The supplied reference already has its config, so rebuilding it is optional.
{
  "id": "new-view", "title": "Your release title", "kind": "panel",
  "paper": "#F7F4EE", "ink": "#17191F", "groundSource": "Your approved palette or sampled app surface",
  "radius": 20, "clearOnExit": true, "primaryTextPx": 24,
  "font": {"family": "ProductSans", "file": "fonts/product.woff2", "weight": "300 900"}
}

primaryTextPx is the measured native CSS size of the decisive result text, not a target invented by the film. paper comes from this product, not the sample above. radius is the component's native radius in CSS pixels. clearOnExit keeps hover until the pointer leaves the real target. Adapt those fields and the optional timing object in config.js; retain the fixed-camera, whole-surface painter in film.js. Native animation frames play at their captured fps. If the action is longer, move the cursor departure and reset later so it finishes and the settled result reads for at least a second; the reference leaves over four seconds. Do not truncate actual processing or portray a queued request as completed.

Keep the resting pointer clear of the critical label. Prefer the visible control itself as rects.target; for a labeled row, clip.json may set "targetFraction": [0.72, 0.5] to choose a clear point inside that actual hit area (default [0.5, 0.5]). Adapt capture-setup.mjs to move the real mouse to the same fractions of that target's bounding box before down/up. The capture and film must use the same point; moving only the drawn pointer would misrepresent the native press. Check it in decoded phone-width action frames.

Review and deliver

node scripts/render.mjs film --all --check --check-fresh --stress --phone --times key --sheet --review --out film/out
node scripts/export-clips.mjs film

These delivery commands finish. For unattended work, wait for both, complete the review and report the exported film/out/index.html player and media paths. The interactive preview is an attended editing tool: node scripts/preview.mjs film starts a persistent server until stopped with Ctrl-C. Run it separately when someone is there to use it; never append it to an unattended foreground delivery/check command.

export-clips.mjs checks every shape first and refuses new video exports on a failed film check. It then renders silent 30 fps H.264 MP4s, VP9 WebMs and result PNGs into film/out/. Its optional --poster <seconds> selects a settled result (default 3.5). It writes a local player with adjacent context, pause/play and reduced-motion handling. The movie's reset is labeled; it is not a seamless product cycle. Keep the static result poster for reduced motion. The label appears before the reset and remains visible through the end; inspect the decoded last and first frames too.

Read every full-size UI-change strip, hover/press/release and reset in every shape; inspect decoded MP4 frames at those events. Score with critique.md, then get a fresh reviewer if available. Fix the source and repeat only affected checks. No score below 8; after two failed correction rounds, report the remaining fault rather than hide it. The stress check covers drawn words, so inspect captured native text separately at 358 px.

Deliver editable films with all local captures/fonts and licenses, MP4/WebM/poster paths, release captions, source claims, the regeneration command and REVIEW.md with scores and limitations. Open the finished MP4s for review. Aim for a web clip under 1 MB, preserving legibility first; report actual bytes and elapsed time.

The default loops are silent. music/README.md explains the included optional song made with AI (selected take A). Never substitute a synthesized score. A supplied or selected song uses the house fitter; keep the silent exports for embeds.

Invariants

  • Real product captures and the user's own/sample data only. No invented UI, lifted rectangles, text patches, crossfades between states or fake processing.
  • Complete native enclosure, even opposite padding, fixed scale and camera. No additional click ripples or device mockups.
  • seek(t, values) paints from arguments only; ctx.reset() every frame, CPU-backed main/helper canvases, prepared local images. No clocks, random calls, timers or network in the painter.
  • Every film word is a template.fields entry. Critical native text must read independently of drawn-copy stress checks.
  • A reset never pretends to undo a purchase, send or irreversible action. Use a sample-state reset explicitly labeled, or play once and retain the real result.

Ostlery is a fictional company made for omnigami's examples.

This skill is omnigami's, © 2026 Auberon LLC. Use what it makes for anything, commercial work included, and change the skill for your own use; don't redistribute or sell the skill. Full terms: LICENSE.md

Individual skills in this repo

This repo contains 4 individual skills — each has its own dedicated page.

Related Skills