← ClaudeAtlas

clip-raw-footagelisted

Cut raw talking-head footage (phone, camera, Zoom) into clean, transcribed, sentence-bounded clips with retakes flagged, ready to drop into a talking-head ad cut. Use when someone hands over raw takes and wants usable pieces.
wrg32786/aigent-talking-head · ★ 3 · AI & Automation · score 71
Install: claude install-skill wrg32786/aigent-talking-head
# Clip raw footage `kit/clip.py` turns one raw recording, or several takes of the same script (usually phone clips from `inbox/`), into numbered clips you can pick from, in four steps (five with several takes). On Windows run it as `py -3 kit/clip.py`. If the goal is a finished reel, `npm run reel` runs this for you; see the `talking-head-ad` skill. 1. **Transcribe.** It runs faster-whisper with word timestamps over the whole file. 2. **Split into sentences.** Short sentences merge with a neighbour until each clip is at least 3 s long, because shorter inserts read as jumpy. 3. **Flag retakes.** If a later sentence within 150 s repeats any 4-word phrase, the earlier one is marked `retake (see clip-NNN)`. A short lead-in that runs straight into a flub ("All right, welcome back.") gets the same mark. Restarts often reword the opener but keep the body, which is why it compares phrases rather than first words. **Nothing is deleted.** 4. **With several takes, pick the best take of each sentence.** Sentences are matched across files by a shared 4-word phrase or by wording that's at least 60% the same. Each match is scored on Whisper's word confidence, minus repeated words and fillers ("um", "like"), and the best one stays `keep`; the others are marked `retake (take2.mp4 is better)`. `--hook take3.mov` forces the first sentence from that take, for example the one shot from your most extreme angle (optional). 5. **Set clip ends from the audio, not from word stamps.** Each clip start