Claude Code video editing skill — automatic rough cut of your raw footage

Let Claude edit your talking-head videos. You hand it the raw clips, it
gives you back the rough cut: silences removed, false starts removed,
only the last take of each sentence kept, no word cut in half, plus
the transcript. One script, a handful of settings, nothing else. Runs
on your own computer (macOS, Windows, Linux), nothing is uploaded
anywhere, no account, no API key.

you:     edit my video, the clips are in Downloads
claude:  joined footage: 538.0s, 1 clip(s), 29.97 fps
         islands of sound: 71 (silences >= 0.22s at -32 dB)
           cut   24.17 ->   25.09  retake  "La plupart des gens..."
           cut  250.48 ->  253.03  retake  "Si tu as encore de plus en plus d'utilisateurs sur ton application,"
           cut  254.11 ->  255.63  retake  "si tu as encore de plus en plus"
           ...
         kept 54 segments: 538.0s -> 117.8s (22 retakes and 3 islands without speech dropped)
         -> edited.mp4, edited.txt, edited.srt, edited.takes.tsv

A 9 min raw take became a 1 min 58 video, in about two minutes
of compute on a Mac.

Install

1. Put this folder in your Claude Code skills:

   • macOS / Linux: ~/.claude/skills/video-editing/
   • Windows: C:\Users\<you>\.claude\skills\video-editing\

   git clone https://github.com/vincentventalon/claude-code-video-editing-skill ~/.claude/skills/video-editing
   

2. Open Claude Code and say: "edit my video, the clips are in Downloads".

That's it. If a tool is missing (ffmpeg, whisper), Claude installs it: the
skill tells it how. The transcription model downloads itself the first time.

Which model for your computer

medium is not worth it: as heavy as the default and slower.

What you get

edited.mp4 (the rough cut), edited.txt (the text), edited.srt
(subtitles with timings) and edited.takes.tsv (every piece of your
footage, kept or cut, with what you said in it).

Check it, cut what is left, leave notes

python scripts/review.py edited.mp4

A page opens in your browser: the video, what you say sentence by sentence,
the sound wave with every join, and what the script already cut. You do two
things there. Cut a sentence it missed (click it, press X: you hear the
cut straight away). Leave a note on a moment (N: "too long here", "keep
the first take"). It saves as you go. Then tell Claude "apply my edits": it
cuts, reads your notes and does what they say. No install: it is Python and
a single HTML file.

Finish it with Remotion (optional)

If you have Node, remotion/ turns the rough cut into a reel: captions
where the word being said lights up, a title on top. Two layouts, no split
screen: you full screen, or your video in a card on a dark
background, for a clip filmed sideways. It picks by itself from the way you
filmed. Ask Claude "finish it with Remotion".

How it cuts

1. Joins your clips, in the order you give them, without re-encoding.
2. The sound decides where to cut. The audio is split into islands of
   sound, whatever lies between two silences. A cut can only fall in a
   silence: never in the middle of a word.
3. Each island is transcribed on its own (whisper.cpp, locally).
   Transcribed in one go, whisper glues two takes into one sentence and
   quietly writes the repeat only once; island by island, it cannot. An
   island with no word in it (a breath, a click, a lone "um") goes.
4. The last take wins. When an island is the beginning of what you say
   right after, it was a false start: it goes, the last take stays. Every
   island is also re-split on its short inner pauses and each piece is read
   again on its own: that is where the "and so you, and so now you have"
   that whisper had swallowed come out. Shoot the way you talk: say it
   again until it is right, and move on.
5. Renders an .mp4 ready to post, picture and sound in sync to the
   frame, a 10 ms fade at each join so you never hear a click.
6. Claude reads the takes like an editor: a sentence you abandoned and
   restarted in other words is something only a reader sees. It cuts it
   with one more command, which re-renders without re-transcribing.

What it does not do (yet)

Without Remotion: no burned-in subtitles. And no cover on frame 0, no music, no b-roll. I have
all of that in my own pipeline; if you want one of them here, open an issue
or say so under the video, and I will add the ones people ask for.

Why it is so small

I post one video a day, edited by Claude, since August 2026. My real
pipeline is 1,400 lines of instructions and thirty scripts: word-level
subtitles, title cards, covers, a TikTok and an Instagram render, automatic
posting. It is far too personal to be useful to anyone else as is.

This repo is its cutting engine, simplified: get the audio, remove the
dead air and the retakes. It works right away, and it is meant for you to
iterate on with Claude, the way you shoot.

The next steps are yours

Ask Claude, one video at a time, in the order you miss them: "burn the
subtitles into the video", "put a title on the first three seconds", "add
music under my voice". It edits this script. That is exactly how my version
grew.

Settings

--lang fr for French (or any whisper language), --cut 12.5:14.0 to
remove a window and --protect 12.5:14.0 to keep one (times of the joined
footage, as in edited.takes.tsv), --pad 0.08 for more air around each
island, --keep-retakes / --keep-fillers to leave them in,
--no-transcript to skip whisper (silences only).

For the maintainer

dev/ is not part of the skill. dev/check_edit.py measures an edit
(duration, A/V desync, words cut at the joins, repeats left).
dev/sync_from_private.py carries the cutting improvements of my private
pipeline over to this repo and tests them on a real rush — read its header.
It never commits nor pushes.

🇫🇷 En français

Tu clones le dossier dans tes skills Claude Code (~/.claude/skills/video-editing/
sur Mac, C:\Users\<toi>\.claude\skills\video-editing\ sur Windows), tu
ouvres Claude Code et tu lui dis « monte ma vidéo, les rushes sont dans
Téléchargements ». S'il manque un outil, c'est lui qui l'installe. Il
coupe les blancs, les faux départs et les reprises (il garde toujours ta
dernière prise), sans jamais couper un mot en deux, et te rend la
transcription. Tout tourne sur ta machine, sans compte ni clé d'API.

Ensuite, « ouvre la page de relecture » : tu vois la vidéo, le texte et
l'onde sonore. Tu coupes ce qui reste (clic sur la phrase, X) et tu laisses
des notes (N). Puis « applique mes coupes ». Tu as Node ? « finis-la
avec Remotion » : sous-titres et titre, toi en plein écran ou ta vidéo
dans un cadre si tu as filmé à l'horizontale.

Ordinateur modeste ? Le modèle par défaut pèse 1,6 Go. Sur un PC sans
carte graphique, dis-lui « prends le modèle small » (0,49 Go, plusieurs
fois plus rapide, rate un peu plus de reprises) ; si c'est la place ou la
mémoire qui manque (8 Go), « prends large-v3-turbo-q5_0 » (0,57 Go,
presque aussi bon). medium ne vaut pas le coup : aussi lourd, plus lent.

Vidéo, contexte et la version française de cette page : deviensdev.fr/claude-code-montage.

License

MIT — Vincent Ventalon.