/video-edit
Inputs: $0 slug, footage file(s) copied into projects/<slug>/assets/footage/ (on the user's computer via the device bridge, or uploads). Optional: brief, notes with timestamps.
Outputs: renders/<slug>_vN.mp4, EDIT.md (paper edit + EDL), captions (assets/captions.srt), review files.
Load only
docs/editing-and-compositing.md§ Editing original footage, § Audiostyles/talking-head-editorial.mddocs/quality-assurance.mdbefore reviewing
Procedure
- Ingest:
python tools/ingest_reference.py $0 assets/footage/<file> --name raw --transcribegives shots, keyframes, loudness and a word-timed transcript (needspip install faster-whisper). Inspectrefs/raw/contact.png. - Paper edit (EDIT.md): list phrases with
[in–out]. Mark retakes (keep the last clean one), false starts, filler and long pauses. Cut-safety rules: gaps ≥400 ms are clean cut points; 150–400 ms only after checking frames; never cut inside <150 ms; pad 30–200 ms; keep thought-pauses ≤0.35 s. Never reorder words in a way that changes a claim. Never speed up speech. - The user's voice is the narration. Clean it: high-pass 80 Hz, compression 2.5:1, loudness to the master. Don't replace it with TTS unless asked.
- Graphics where they help: define a term, show a number (sourced), draw a mechanism, highlight a screen region. Place them in the opposite third from the face, or cut to full-screen graphics while the voice continues. Never cover faces or important screen content (check the face box in keyframes).
- Build
engine/src/projects/<slug>/Scene.tsxwith an EDL array[{src, in, out}]→<OffthreadVideo>sequences, plus graphic overlays anchored to transcript words (put the transcript segments in timeline.jsonnarration, then cues as usual). - Captions: word-timed, 1–2 lines, ≤ 32 characters per line, inside safe areas, never duplicating on-screen kinetic type.
- Mix (
audio_mix.pywith the cleaned dialogue asaudio/narration.wav), render,/video-review, revise. - Frame-specific feedback: when the user says "at 0:12 the graphic covers my hand", extract that frame, fix it, re-render, and show the before/after still.
Quality gates
- Meaning preserved: diff the kept transcript vs the original; no altered claims
- No jump-cut artefacts flagged as pops; no faces covered (check stills)
- Footage, music and images used are owned or licensed (manifest)
- Loudness −14 LUFS, dialogue intelligible (measure; ask the user to listen)