I made Claude Code my video editor (cuts, captions, exports)

AI video editing automation sounds fancy, but an edit is really four mechanical moves: transcribe, cut the dead air, caption, export. Claude Code can run all four with free tools, and a skill file means you teach it your taste exactly once.

What you'll have when you're done

• A one-command pipeline: raw clip in, cut and captioned 9:16 export out
• A skill file that holds your editing rules, so every future session already knows them
• A cut list you approve before anything renders, so you stay the editor in chief
• Zero per-video costs and no watermarks
• [SCREENSHOT: before/after, raw clip vs captioned 9:16 export]

Before you start

• Claude Code installed (it's Claude in a terminal that can run tools on your machine). Comes with the Claude Pro plan, $20/month.
• ffmpeg and Whisper, both free. Don't install them by hand; Claude Code does that in step 1.
• A raw talking-head clip to practice on, ideally 1-3 minutes with some obvious dead air.
• 30 minutes.

Step 1: Let Claude Code set up the tools

Open Claude Code in a new folder (call it editor) with your practice clip copied inside. First prompt:

Install ffmpeg and OpenAI's Whisper (the open-source transcription
tool) on this machine if they're missing, the standard way for my
OS. Then verify both: run ffmpeg -version and transcribe nothing with
whisper --help. Tell me when both respond.

ffmpeg is the command-line video tool almost every video app is secretly built on. Whisper is a free transcription model that runs locally, no API key needed. Claude Code handles the install commands and tells you if anything needs your password.

Step 2: Write the editor skill file

A skill file is a saved instruction document Claude Code reads before working, the same idea as a voice file but for editing taste. Create it once:

Create a file called EDITOR.md in this folder with exactly this
content, then confirm you'll follow it whenever I ask for an edit:

You are my video editor. When I give you a clip, run this pipeline:

1. TRANSCRIBE: run Whisper on the clip. Save the transcript with
   word-level timestamps as [clipname].json and a readable
   [clipname].srt next to it.
2. CUT: find silences longer than 0.8 seconds using ffmpeg's
   silencedetect filter. Cut them out, snapping every cut to the
   nearest word boundary from the transcript. Never cut mid-word.
3. CAPTIONS: burn the transcript in as captions. Style: uppercase,
   max 4 words per line, centered in the lower third, white text
   with a thick black outline.
4. EXPORT: 1080x1920 (9:16), H.264, keep the source frame rate,
   audio untouched.

Rules:
- Never modify the original file. Every output is a new file next
  to it: [clipname]-cut.mp4, then [clipname]-final.mp4.
- Show me the cut list (timestamps in and out, total time removed)
  and WAIT for my approval before rendering anything.
- If the audio clips or peaks, flag it, don't fix it silently.

Step 3: Run the first edit, approve the cut list

Now the whole pipeline is one sentence:

Read EDITOR.md, then edit myclip.mp4.

Claude Code transcribes, finds the silences, and stops to show you the cut list: every cut's in and out point plus the total dead air removed. This approval gate is the difference between an editor and a slot machine. Scan the list; if a "silence" is actually a dramatic pause you want kept, say so and it re-plans.

Step 4: Render and review

Say "approved" and let it render. Two files appear: the silence-cut version, then the final with captions burned in at 9:16.

Watch the final top to bottom once. Check the three usual suspects:

• Cuts that clip the first syllable of a word (tighten the 0.8s threshold to 1.0s if so)
• Captions drifting out of sync in the second half
• The 9:16 crop cutting off your face (tell it where to anchor the crop)

Anything wrong goes back as one note, and the fix goes into EDITOR.md, not just this edit:

The cuts feel too aggressive, raise the silence threshold to 1.0
seconds. Update EDITOR.md so this is permanent, then re-run the edit.

Step 5: Make it the default workflow

From now on, editing is: drop the raw clip in the folder, open Claude Code, one sentence. The skill file carries your taste between sessions, and each calibration note makes the next edit closer to hands-free.