# 04 · Claude Video Editor

You film, drop the clips into a folder, and tell Claude to edit. It picks the best take of every line, cuts the
video, levels the sound, tracks your face so captions don't cover it, adds captions in your font, a title card
and grain, and hands back the finished video with a transcript, captions and a check file.

**For your AI:** walk me through building this one step at a time. Ask how I film and what my videos look like
first. Don't move on until the current step works.

## First, check

- Do you film talking videos every week, or pay an editor for them?
- Are your videos mostly you talking to camera? This works for that. It doesn't replace an editor for b-roll stories.

## What you need

- A Mac or PC. ffmpeg does the cutting, sound levels, captions and grain.
- A transcription service that gives a time for every word (AssemblyAI, Deepgram or Whisper). Only the
  audio has to leave your computer, and it costs cents per video.
- Claude Code or another AI that can read files and run commands on your computer.
- Your font and your rules: where captions sit, how big, how many words at once, the look.

## Where it runs

On your computer. One folder in, one folder out.

## The steps

1. **The folder.** A "Drop clips here" folder and a "Finished" folder. One command starts an edit.
2. **Transcribe.** Send the audio for transcription and keep every word with its start and end time.
3. **The take map.** Your AI reads the transcript, finds lines you said more than once, and picks the best
   take of each. It writes the cut as a list of start and end times, never by hand in an editor.
4. **Cut and level.** ffmpeg cuts the list into one video and brings the sound to one level (about -14 LUFS
   for social video).
5. **Captions.** Captions from the same word times, in your font. Track your face so the words never sit on it.
6. **Finish.** Title card, grain, the shape you post in (vertical 9:16).
7. **The check file.** Measure the result: length, sound level, caption timing, words off the face, no
   pauses left. A check you can read in ten seconds beats watching it three times.

## Checklist before you rely on it

- [ ] Your originals are never changed.
- [ ] The captions match what you said (read the transcript once; names and numbers go wrong first).
- [ ] The check file fails loudly when something is off, so plant a mistake once and see it go red.
- [ ] You watch the finished video once before posting. Always.
- [ ] Your rules (font, size, position, look) are written down, so every video looks the same.