Video Editing, Captions, and Transcripts

Critical rules

Choose the mode first:
- Editing requested: read references/workflow.md + references/decision-rules.md.
- Captions/transcript only: read references/captions.md + references/deliverables.md; do not run the editing pipeline.
Captions: prefer stt if on PATH; else ask user (suggest install stt / existing SRT / ElevenLabs). Generate from the source once. Never silently switch providers.
Ship: follow references/deliverables.md. Editing defaults to 2K mp4 + SRT; caption/transcript-only jobs deliver only requested text/timed-text files.
Face cam: read references/face-pip.md before any PiP work.
Glitches: read references/failure-modes.md when output has caption, join, render, sync, or crop problems.
For editing, source of truth is timed speech (SRT), not silence detection.
Keep successful performances, including clean attacks/releases. Never leave prep-breath / mouth-start into a deleted attempt (out: surgical).
Use bundled scripts/. Re-encode cuts. Put all temps under work/.