Changelog → Branded Video

Input: a changelog .md (themes + items, like the weekly HyperFrames digest). Output: a lint-clean, seam-gate-green HyperFrames project in projects/active/weekly-changelog-<range>/. Render only when asked.

Load first, non-negotiable: motion-doctrine (+ cut-the-curve, oversized-cursor if a cursor appears, seam-craft) and captions-overlay. This skill supplies the changelog-specific pipeline; the doctrine supplies the motion law.

The prime directive: visualize, don't list

Every theme is illustrated by an animated mock of the actual UI or a faithful analog acting out the change in experience — never text bullets. Route every theme/item through references/visualization-registry.md BEFORE writing the script; the registry decides ui-recreate / ui-analog / terminal / checklist. Text checklist is the LAST resort, reserved for genuinely non-visual items (reliability fix lists).

Pipeline

0 · Bootstrap the project from THIS skill's assets — non-negotiable

Do this before writing any composition HTML. Skipping it always produces a video that looks like a similar project you built before, NOT this skill's brand — that's the single most common way this skill goes off-brand. The skill's assets, fonts, and scaffold are the skill; the SKILL.md prompt is a router.

mkdir -p project/assets/fonts
cp <SKILL_DIR>/assets/fonts/*.woff2 project/assets/fonts/
cp <SKILL_DIR>/assets/bgm.mp3 project/bgm.mp3
ffmpeg -y -stream_loop 15 -i <SKILL_DIR>/assets/bg-pattern.mp4 -t <TOTAL> \
  -vf "scale=1080:1080,fps=30,eq=saturation=0.72,drawbox=c=black@0.5:t=fill" \
  -an -c:v libx264 -crf 20 -pix_fmt yuv420p project/assets/bg-pattern-<TOTAL>s.mp4
cp <SKILL_DIR>/examples/master-skeleton.html project/index.html

Then read references/build-spec.md end-to-end (not skimmed) — it defines the brand tokens (TT Norms Pro + ABC Solar Display + TT Norms Mono, cream #f5f6f4, rationed green #5ef17c, glass cards with green-tinted borders, kicker/sec-chip pill shape, 32px caption rail at top: 990) that every scene inherits from the scaffold.

Only THEN begin steps 1-6 below. Steps 1-4 (parse, route, script, VO) plan what goes into the scaffold; step 5 fills placeholders (<RANGE>, <TOTAL>, <CUT_N>, <DUR_N>, scene bodies) inside the already-copied project/index.html — you do NOT rewrite the scaffold's chrome, fonts, palette, or layout shell.

If you catch yourself reaching for cp on a prior video's index.html, or writing your own @font-face declarations, or designing a WebGL shader background instead of using the encoded bg-pattern MP4 above: STOP. Delete the current index.html and restart at the cp of the master-skeleton scaffold. Rebuilding scene content on the right scaffold is cheaper than retrofitting brand into the wrong scaffold.

1 · Parse + editorial cut

• Extract: week range, headline stats (releases, commits), themes, items.
• Budget: 45-60s total. Title ≤2s, outro ≤3.5s, 4 themes ≈ 9-12s each.
• Per theme keep ONE hero visualization + at most 3 spoken items. Everything else exists only as the outro's "full digest" pointer. Cutting is the job: a changelog with 30 items still yields ≤14 spoken beats.
• Order themes by story: marquee feature → product surface → performance → reliability (the digest usually already reads this way).

2 · Visualization routing

For each theme, pick the surface from references/visualization-registry.md and write one line: theme → surface → the 2-4 sequenced actions the mock performs, each tied to a script phrase. If no registry surface fits and no faithful analog exists, it's a checklist scene — don't invent fake UI for something we can't represent honestly.

3 · Two-layer script (spoken vs display)

Write the script as token lines per references/script-voice.md: conversational register, every technical term carrying a spoken phonetic form from references/lexicon.json while display keeps standard spelling. Captions show display; the VO reads spoken. Any term not in the lexicon: STOP and ask the user how it's pronounced, then add it to the lexicon. Save as script-tokens.json in the project.

4 · VO — Annie (HeyGen, pinned)

# spoken-layer text only; words JSON = ground-truth timestamps of the SPOKEN text
# Repo-native path: the changelog-video skill runs from the hyperframes repo root,
# so it uses the tracked hyperframes-media TTS helper directly (no `npx hyperframes
# skills` install step). If you've copied the skill into another repo, swap in
# your own path to the media-use / hyperframes-media heygen-tts.mjs.
node skills/hyperframes-media/scripts/heygen-tts.mjs ./vo-spoken.txt \
  -o voiceover.mp3 --words vo-words.json \
  --voice 330290724a1b470fb63153f34d4c0183   # Annie — lifelike (do not substitute)

Requires heygen CLI ≥0.3.0 authenticated (heygen auth login --oauth). Then align spoken timestamps back to display tokens:

node <SKILL_DIR>/scripts/align-captions.mjs \
  --tokens script-tokens.json --words vo-words.json --out captions.json

captions.json is the caption-rail input (display spelling, spoken timing). The aligner prints MISMATCH warnings — resolve every one before building (usually a lexicon spelling the TTS renders as multiple words). The audio is the clock: all beat times come from vo-words.json; a VO regen re-opens every seam.

Word-timings are a hard gate. Before moving on to step 5, verify vo-words.json is non-empty and has a words: [...] array with start/end per word. If it's empty (0 bytes) or missing the array — a known failure mode when the TTS provider returns audio but no timestamp payload — DO NOT proceed without them. Fallback: forced-align the produced audio against the display script using local whisper:

uvx --from openai-whisper whisper voiceover.mp3 \
  --model base.en --language en --word_timestamps True \
  --output_format json --output_dir .
# then run align-captions.mjs with --words voiceover.json (same shape)

Whisper mishears TTS renderings ("gee-sap" → "gsap", "heyjen" → "hey Jen", etc.) — captions still use the DISPLAY spelling from script-tokens.json; whisper only supplies the timestamps. align-captions.mjs handles the join. This fallback is the difference between a captioned build and a silently uncaptioned one.

5 · Build

Follow references/build-spec.md exactly: brand tokens + fonts (bundled in <SKILL_DIR>/assets/), the animated background encode, scene scaffold, chrome, caption rail, one rationed green moment per scene. Then the doctrine order: ledger.json (all ordinary seams cut-the-curve LEFT) → seam-stamp → internal beats on VO words → seam-gate verify.

Captions are non-optional. The master-skeleton ships a caption-rail IIFE that reads a LINES array — leaving that array empty is a shipped bug, not a style choice. Populate it from captions.json before proceeding to step 6:

// paste in place of "const LINES = /* … */ []" in the caption-rail IIFE:
const LINES = /* contents of captions.json */ [
  { id: 0, end: 2.74, w: [["This", 0.0], ["week,", 0.30], …] },
  …
];

If align-captions.mjs was skipped or LINES is [], the frame check in step 6 will fail — do not paper over it by removing #cap-line from the scaffold.

6 · Gates (all green before presenting)

1. bun run --cwd packages/cli hyperframes check --caption-zone "x0=0;y0=.90;x1=1;y1=1;severity=error;seek=.02,.06,.10,.14,.18,.22,.26,.30,.34,.38,.42,.46,.50,.54,.58,.62,.66,.70,.74,.78,.82,.86,.90,.94,.98" (or the installed hyperframes CLI from the repo-local skills/hyperframes-cli/ skill) — 0 errors (contrast: dim text ≥ .66 alpha; scene content stays above the caption rail). Do NOT reach for npx hyperframes@latest; the tracked repo-local CLI is the source of truth for the composition contract this skill produces against.
2. seam-gate.mjs verify — 0 fail.
3. Restart the preview server (it caches the bundle), spot-check 3-4 beats via __player.seek on the raw comp page.
4. Do NOT render unless the user asks. After a requested render, verify frames from the MP4 (ffmpeg -ss <t> … -frames:v 1): captions present, background video not black, no tiny/frozen frames.
5. Caption presence gate — hard fail. Sample 3-4 frames spread across the VO's spoken window (e.g. t=3, t=15, t=30, t=42 for a 48s VO) and confirm the caption rail at top: 990 renders visible text on each. If any frame in a spoken interval is missing captions, the build ships uncaptioned — treat it as a red gate and re-check step 5's LINES population. This is exactly what went wrong on the Jul 13-20 v4 build.

Project layout

projects/active/weekly-changelog-<range>/
├── index.html            # single-doc master (scenes as slides, stamped seams)
├── ledger.json           # vector ledger (seam-stamp input)
├── script-tokens.json    # two-layer script (source of truth for VO + captions)
├── vo-spoken.txt         # generated: spoken layer, one line
├── voiceover.mp3 + vo-words.json + captions.json
├── bgm.mp3               # copy from <SKILL_DIR>/assets/bgm.mp3 (the house track) unless the user supplies one
└── assets/fonts/ + assets/bg-pattern-<dur>s.mp4

Anti-patterns