Long-form editor

Everything is code, so every change is repeatable: a note like "music's too loud" becomes one number in one script.

Folder layout

<video>/raw/          IMG_xxxx.MOV takes (+ screen recordings)
<video>/editor/       these scripts (copy from scripts/)
<video>/final/        output + upload text

Run everything from editor/. Needs ffmpeg, Python (numpy scipy soundfile pillow faster-whisper), Node + playwright (npm i playwright && npx playwright install chromium).

Steps

1. Check the footage first with the video-quality-check skill (HDR? full range? dim?). Fix TM/GR in build.py to match.
2. Transcribe. Extract 16 kHz mono to tr/IMG_x.wav and 48 kHz stereo to a48/IMG_x.wav, then py tr.py IMG_x ... (Whisper with word timestamps).
3. Find the story. Read every transcript. Write segs.py: (section, take, first-word start, START of last word[, hard out[, hard in]]). Chapter slots are ('CARDn', None, seconds).
   • Hook first: the result, then the promise. Cut stammers and repeated phrases; prefer the clean re-say.
   • Re-transcribe a 5–10 s clip whenever a timing looks wrong. Whisper merges repeats and invents words in silence.
4. py edl.py → plan.json. Cuts snap to the quietest 10 ms near word edges, and pauses over 0.6 s shrink to 0.3 s.
5. py words.py → word times on the output timeline (g/words.js).
6. py build.py → graded, range-converted, punch-in A-roll (aroll.mkv). Set the face position per take in FACE so zooms stay centred.
7. py voice.py → voice.wav (crossfaded joins, EQ, de-ess, gentle compression, ~-14.5 LUFS).
8. QA the cut: re-transcribe voice.wav and read it start to finish. Fix segs.py and repeat 4–8 until it reads clean.
9. Graphics: write g/_content.js (see example_content.js), assemble part2.js = _infra.js + _content.js + _tail.js, then node check.mjs.
   • block(section, 'start phrase', 'end phrase', html, {side}). Children with data-at="phrase" reveal on that word.
   • One graphic at a time, opposite side from the face, never over it. Render sample frames and look at them before the full render.
10. py music.py → original score + sound effects timed to the sections and graphics (no licensed music needed). Then py mix.py → music ~20 dB under the voice, oversampled true-peak limiter.
11. Render graphics: in g/, run node render.mjs <from> <to> in 3–4 parallel chunks.
12. py final.py → one composite pass, CRF 14 slow, AAC 320k.

Taste rules

• One focal point on screen. No subtitles top and bottom at the same time.
• Graphics say less than the voice: 2–4 words big, one small line under.
• Never put a number on screen that wasn't said or verified.
• Chapter cards between steps; a short end card.