Cut raw talking-head footage into a finished edit with video-use
Claude Code cuts the filler and dead air from the transcript, adds subtitles and simple grading, checks its own render, and hands back a clean edit to review.
The problem
A ten-minute recording takes an hour or more to edit by hand: removing ums and false starts, closing gaps, choosing takes, and adding subtitles.
What you get
A working system with the steps, tools, checkpoints, and expected return made explicit.
- Setup
- 40 minutes
- Back each week
- 3 hours
- Difficulty
- assisted
Expected return
The working case
A planning estimate, not a guaranteed result. Measure the first four weeks against your own baseline.
144
hours returned per year
At 3 hours/week across 48 working weeks.
1
week to earn back setup
Compare the setup estimate with the weekly time returned.
Operating contract
Input
The source material, constraints, and examples a human would need to do this work well.
Checkpoint
A person reviews judgment calls, sensitive content, unfamiliar tools, and irreversible actions.
Success signal
Track time returned, corrections required, and exceptions. Keep it only if the measured result compounds.
Before you start
- ·Claude Code
- ·ffmpeg, yt-dlp, and uv installed
- ·An ElevenLabs API key
The steps
- 01
Install video-use as a Claude Code skill. Clone https://github.com/browser-use/video-use, link it into ~/.claude/skills/video-use, run uv sync, install ffmpeg and yt-dlp, and copy .env.example to .env. Or ask Claude Code: "Set up https://github.com/browser-use/video-use for me."
- 02
Add your ElevenLabs API key to .env. video-use sends the audio there for a word-level transcript.
- 03
Put the raw takes in one folder and tell Claude Code what the final video is for.
Copy this prompt
Use the video-use skill on the files in [folder]. This is a [length] [purpose] video for [channel]. Cut filler words, false starts, and dead air. Keep the best take of any repeated line. Add burned-in subtitles and a light color grade. Show me the cut list before you render.
- 04
Read the cut list, fix any cut that changes the meaning, then let it render and self-check the output.
- 05
Watch the full render once at normal speed before you publish it.
What it runs on
- video-use
Open source Claude Code skill from Browser Use. Edits from the transcript with ffmpeg.
- ElevenLabs
Transcription. Needs an API key and bills per audio minute.
- Claude Code
Runs the skill on your machine.
Where this goes wrong
- The model reads the transcript, not the picture. It will not notice a blink, a bad frame, or someone walking into the shot.
- Tight cuts on every pause can sound rushed. Tell it how much breathing room to keep.
- Your audio leaves the machine for transcription. Do not use it on footage you may not share with a third party.
Definition of done
Run it for four weeks. Then make it earn its place.
- □ Baseline the manual time before launch.
- □ Keep a human approval step for consequential output.
- □ Record corrections and exceptions, not just successes.
- □ Expand, revise, or retire it after the first review.
Build the system around it
Related workflows
If this one stops working, tell us. Three reports in a month and it leaves the library until a person has looked at it again.