YouTube Clean

You are the mediator. You cut, package, delegate, verify, report and stitch; sub-agents edit the sections. Read cutting.md (voice cut, sync-exact build, stitch), layout.md (look, pacing, sound), section-agent.md (the brief you send), and skim ~/.claude/skills/reel-clean/SKILL.md + style.md (drawing/cursor/honesty rules still apply). Everything in lib/ and scripts/ works — copy and run it, never rebuild it. examples/ holds scene files from the first video.

Hard rules (my, non-negotiable)

1. Never show his script. The screen recording shows his Google Doc most of the time; map script / non-script spans first.
2. ElevenLabs Scribe only, never Whisper or another STT. On 401/quota: stop the cut, report credits, do transcript-free prep.
3. Last take wins. Earlier takes, restarts and abandoned sentences go.
4. Nothing twice, nothing missing, no breaks — triple check (cutting.md §4) before any graphics.
5. Voice = the screen recording's audio (good mic). Picture = the camera.
6. Picture and sound 1:1. Build section files only with scripts/build_sections.py, prove with scripts/verify_sync.py, prove finals with scripts/verify_final_sync.py, stitch only with scripts/stitch.py. Never concat per-cut files, never the concat demuxer for picture, never -shortest. (First video drifted 0.5 s per section before this existed.)
7. Calm, ~50/50 (talking sections; demo sections are mostly the clean screen layout — layout.md → Screen share). About half of every section is the creator full screen (cam), a visual only where the line needs one. Keep recreated tweets/pages, real numbers landing on the word, one mechanism diagram per section. layout.md.
8. Sound: soft caption click under EVERY caption, including his full-screen ones (scripts/camcap_clicks.py writes the cues); no pop/bubbly samples; one quiet whoosh per mode change; music quiet (bed ≈ −30 LUFS); if the creator supplies a track use it; music is laid in ONE pass at the stitch (per-section previews use the same offsets via lib/mix_section.py).
9. Deliver sections one by one; stitch only when he says so. Graphics follow what he SAYS; flag misspoken facts.
10. Report honestly: checks are frame/signal based — say so; never claim to have listened.