produce-video

name

produce-video

description

Use this skill whenever you produce, re-record, or localize a docs tutorial video — a new episode, a new locale of an existing one, a narration or choreography change, a voice/TTS change, or a red `videos` gate in services/docs/tests. It owns the production discipline end to end — storyboard-first authoring under services/platform/tests/docs-videos/episodes/, the billed ElevenLabs TTS stage and its cache, the planned-timeline recorder and its ±100 ms drift gate, the shared demo-org contract, the docs embed set (mp4+vtt+poster ×3 locales), and the watch-it QA gate. Load it before touching services/platform/tests/docs-videos/ or anything under services/docs/public/videos/. Never hand-record, hand-edit, or hand-place a video — every committed asset comes out of `bun run docs:videos`. For still screenshots use write-docs (its SCREENSHOTS.md); for the docs pages around the video use write-docs; for de/fr narration follow write-translations.

Tutorial videos are PRODUCED, not recorded: narration is scripted per locale, the timeline is planned from the measured audio, the browser is choreographed to that plan, and ffmpeg composes the take — so a re-run yields the same video. This file is the discipline; the mechanics live in the pipeline runbook (services/platform/tests/docs-videos/README.md); the docs-page conventions in write-docs; the de/fr voice doctrine in write-translations.

When this applies

Any change under services/platform/tests/docs-videos/** (episodes, recorder, compose, cards), any asset under services/docs/public/videos/**, a <Video> embed on a docs page, a failing services/docs/tests/videos.test.ts, or a request to "make/update/translate a tutorial video".

The rules

• Script before screen. An episode starts as narration text in episodes/<id>/episode.ts — every scene, every locale — and gets a table read (speak it; fix what you stumble on) BEFORE any choreography. The narration is the spine: scene budgets derive from its measured audio, so a script change re-plans the timeline and re-records the take. (Why: pacing edits after recording cost a full re-record anyway — do the cheap iteration on paper.)
• Narration is a colleague showing you, spoken. Contractions, short spoken sentences, action + reason in the same breath, every real choice names a rejected alternative, one pitfall beat and one verify beat per episode, closers recap what the viewer DID — never aphorisms (register rules
  • banned-pattern list: STORYBOARD.md). Written natively per locale (write-translations: du/tu, no bureaucratic German, no marketed French) — never translated word for word. Spoken UI vocabulary matches the shipped catalog (Wissen, Connaissances). Words a locale's voice mis-reads (the brand, product terms) get a respelling in the pipeline's per-locale pronunciation map — spoken text only, captions keep the real spelling. ElevenLabs audio tags sparingly; punctuation and ellipses carry pacing. A wholeTakeLocales locale bills the episode as ONE request — its joined spoken script stays ≤4,500 chars (--stage check guards the 5,000-char cap). Claims must be true of the product — narration ships in three languages and gets quoted back.
• Rehearse free before you bill or record. --stage check (instant: spec ↔ choreography ↔ mock-reply pairing), --stage plan (instant: the timeline table from estimates), then a --mock-tts take (silence-narrated, auto-composed as a draft into .state/) prove the structure, the locators, and the budgets at zero cost. Only then synthesize real narration and record the real take. (Why: every failure mode caught by check/plan/mock costs seconds there and minutes-to-euros later; the same validation runs in bun run check via the episode gate.)
• TTS bills per character — respect the cache. --stage tts is cache-first (.state/tts-cache/); an unchanged scene is free, an edited one re-bills all its locales. Batch script edits, then regenerate once. Voice ids are pinned per locale in the episode spec; audition with --audition before changing one (a voice change re-records every locale). The cache key shape is test-pinned (lib/tts-cache-key.test.ts) — changing it re-bills the entire back catalog.
• Choreography waits on state, never on time — cue() is the one exception. Target elements by role + localeT label or by href/data literal (seeded content stays English in every locale; per-locale DISPLAY NAMES like the builtin agent are data — map them). Scope pickers to their container. A scene must fit its budget in EVERY locale — the recorder throws on overrun; fix with a per-scene minMs/leadInMs in the spec, never by letting it slide. (enforced by the recorder + compose drift gate, ±100 ms)
• Takes are additive-only against the shared org. Recording reuses the docs-screenshots workspace; the only mutation a take may leave is nothing — anything a scene creates on camera registers on ctx.cleanup (thread, knowledge entry, agent, task) the moment it exists, and the recorder sweeps it even on abort. Demo-content changes go through ../docs-screenshots/demo-content.ts and the paired lib/mocks/overrides/docs-replies.ts (the typed prompt must contain its reply's match clause — --stage check enforces the pairing).
• The shipped set is indivisible. An episode ships <id>.<locale>.{mp4,vtt,webp} for ALL three locales, declared in public/videos/manifest.json, embedded via <Video> (src + poster + captions + lang) on pages whose locale matches. (enforced by services/docs/tests/videos.test.ts)

Patterns

Narration-synced beat — the click lands as the voice names it (episodes/ep1-welcome/scenes.ts):

await cue(4.0); // "…we will attach a company document…"
await page.keyboard.type('@', { delay: 60 });
await cursor.click(pickerDoc); // the viewer hears it, then sees it

Before an episode ships — tick every box, or N/A with a reason

• Watched end-to-end at 1×, every locale — not skimmed frames: cursor motion, stream pacing, scene transitions, the fade-out. The per-scene review sheet (.state/review/<id>.<locale>/index.html, written by every compose) is the triage aid, not the watch-through.
• Listened, every locale — pronunciation (brand names!), pace, no synthesis artifacts; spot-check with an STT round-trip when unsure.
• Compose gates green — drift (±100 ms every scene), the automatic A/V verification (duration + speech coverage; never ship on --no-verify), and no overrun waivers.
• Captions open correctly on the rendered docs page in each locale; timing follows the voice.
• bun run --filter @tale/docs test green — the videos contract plus the docs suite.
• Shared org left as seeded — cleanup ran; docs:screenshots --skip-seed --only chat-thread-reply still captures pixel-equivalent.
• Docs pages updated in all three locales (episode page + any embeds) per write-docs.
• Every task block does real work on camera — click/type/submit/create; a hover-only block is context, not a task (the in-depth arc, STORYBOARD.md).
• Pitfall + verify beats present — one real failure shown and diagnosed; the outcome proven on a different surface.
• Table read passed the banned-pattern list in every locale (STORYBOARD.md register), and --stage check is green — no whole-take budget or register warnings left standing.

Companion files

• STORYBOARD.md — read when authoring or restructuring an episode: the storyboard method, scene grammar, and Episode 1 as the worked example.
• services/platform/tests/docs-videos/README.md — read to RUN the pipeline: stack bring-up, stage commands, gotchas.