Skill
Talking-head cut
Give your agent the raw recording of you talking to the camera. It reads the take line by line, finds the false starts, the lines you said again, the notes you left for the editor, the doubled words and the ums, and tells you what it will remove by quoting your words. Every cut lands in a pause, long pauses shrink to a breath, and picture and sound stay in step.
/plugin marketplace add EditAtlas/skills /plugin install talking-head-cut@editatlas-skills
The first line adds EditAtlas skills to Claude Code, once. The second installs this one.
To update, open /plugin, then Marketplaces, editatlas-skills, and choose Update marketplace, then run /reload-plugins. Choose Enable auto-update there to get new versions on their own.
In Claude, open Settings, then Capabilities, and upload the zip under Skills. Code execution needs to be on.
To update, download the zip again and upload it in place of the old one.
npx skills add EditAtlas/skills#v0.22.0 --skill talking-head-cut
For Cursor, Codex, GitHub Copilot and other agents. It asks which of them to install to; add -g to install for every project. Without Node, download the zip and unzip it into your agent's skills folder.
To update, run the same line again; it names the latest release.
Just ask
Cut my mistakes and the dead air out of take.mp4
Clean up this raw take for a Reel, but keep what I meant to say
What it does
- Transcribes the take word by word on your machine and shows it to your agent as numbered lines, with likely mistakes marked.
- Decides what goes: false starts, failed and repeated takes, notes to the editor, doubled words and fillers. Repetition you meant stays.
- Names each removal by its words, never by a time it guessed, and puts every cut in the quiet between words.
- Shortens long pauses to the pace you want and renders picture and sound in one pass.
- Transcribes the cut again to catch anything left in or clipped, and tells you which joins to listen to.
Tested on three raw takes with known answers, in English and Vietnamese: agents with it kept every line meant and removed every mistake in all 6 runs, with picture and sound in step. Agents without it scored 0.57 to 0.86, dropping a repetition the speaker meant and drifting up to 2 frames out of sync. Not yet tested on real recordings.