talking-head-adlisted
Install: claude install-skill wrg32786/aigent-talking-head
# Talking-head ad
The main path: a teammate records on their phone, drops the file in `inbox/`, and you turn it into a reel. It runs in four stages:
- **Clips.** Whisper transcribes the recording (or every take), splits it into sentences, flags flubbed retakes and keeps the best take of each sentence.
- **Cut.** The keep clips are joined with a glitch seam and a 10% zoom punch at each cut, and a bigger eased punch after the hook and into the payoff.
- **Overlay.** A canvas layer animates every spoken word, adds one kinetic title per beat behind the presenter's head, and any cards, styled by a theme and a font pairing.
- **Sound.** A few deep sounds on the key cuts only (locked set). On the effects pass, the 31-sound cinematic pack is auto-picked per effect hit and ducked under the voice; the music bed, if any, also ducks.
Every version gets checked with `npm run verify` and by re-transcribing it. All scripts live in `../../kit`. The rules below come from a working reel workflow; follow them unless the teammate asks otherwise.
## Optional: study a reference, then write the script
Skip this when the teammate already has a script.
1. **Teardown (optional).** If the teammate has a reel they want to match, ask them for the video file and run `npm run teardown -- <file>` (rea-video, Apache-2.0, does the breakdown; see `docs/teardown.md` and the talking-head-teardown skill). It writes a Reel.md with the hook, beats, cut rhythm, title moments, finishes and tempo. Only local fil