Captions

Short-form only. Long-form (16:9) jobs don't get burned-in captions in this pipeline.

LOCK: captions are built from outputs/.transcript.json — the canonical, cut-remapped, corrections-applied transcript — and never by re-transcribing. The embedded-captions vendor skill normally runs its own WhisperX pass; this skill's whole job is to feed it our already- correct transcript instead and confirm it actually skipped its own.

1. Bridge the canonical transcript into embedded-captions' project shape

uv run .claude/skills/captions/scripts/prepare_captions_project.py <job-name> [--language en]

This creates projects/<job>/captions/ with:

• source.mp4 — a copy of the latest render (outputs/<job>-graphics.mp4 if graphics are done, else outputs/.roughcut.mp4)
• transcript.json — outputs/.transcript.json converted into embedded-captions' exact normalized schema ({"words":[{"text","start","end","type":"word"}], "language_code"}).

If this job's audio isn't English, pass --language <code> — the default is en.

2. Run embedded-captions against projects/<job>/captions/

Follow /embedded-captions from here, using projects/<job>/captions/ as its project directory. Its own scripts/prepare.sh runs matting, transcription, and audio-envelope in parallel — watch its log for the transcription step and confirm it reports finding an existing transcript.json and skips re-running Whisper. If it re-transcribes anyway (schema drift, a version bump changed its expectations), STOP — do not let it overwrite anything in this pipeline's transcript/ or outputs/.transcript.json, and fix prepare_captions_project.py's output schema instead of accepting a fresh transcription.

Pick the caption identity/style per the user's ask — the quiet anchor rail is the default; only go fully cinematic (embedding words behind the subject) if the user explicitly wants that treatment.

3. Safe zones still apply

Whatever embedded-captions produces, the burned-in captions must land inside y 200–1620 on the 1080x1920 canvas — no key visuals (including caption text) in the top 200px or bottom 300px.

4. Bring the captioned output back into this job

Copy embedded-captions' finished render into projects/<job>/outputs/<job>-captioned.mp4. Continue to /music (optional) or straight to /export.