Can Claude edit videos?

Yes, through Claude Code, not the chat window. Claude can't watch video, but it can read a word-level transcript and a few still frames, decide the cuts and run FFmpeg to make them. That covers silence and filler removal, captions, motion-graphics overlays and a 9:16 export; footage without speech, color and taste still need you.

How Claude edits a video it can't watch

Claude models read text and images, not video, per Anthropic's model overview. Every working setup uses the same trick: turn the footage into something Claude can read. A transcript with a timestamp on every word tells it what was said and when. A few still frames, pulled only where a decision needs them, tell it what's on screen. The video-use README describes it exactly that way: the model reads the video instead of watching it.

From there, Claude writes an edit decision list, a plain file of in and out points, and runs FFmpeg to cut, join, caption and export. That's why this lives in Claude Code and not in a regular chat: the agent has to run commands on your files.

Timing is everything, so use word-level timestamps. OpenAI's Whisper is free and runs locally, but the WhisperX README notes that Whisper's timestamps are per utterance and can be off by several seconds. WhisperX adds forced alignment for word-level timing and can label speakers, which is what you want when a cut has to land between two words.

What to install

You need Claude Code (on a Pro, Max, Team, Enterprise or Console account), FFmpeg, Python and a transcriber. auto-editor is optional but handy for the silence pass; it's no longer published on pip, so use Homebrew or the release binary. On a Mac:

curl -fsSL https://claude.ai/install.sh | bash
brew install ffmpeg auto-editor
# Captions on a Mac: the standard ffmpeg has no libass, so add ffmpeg-full (keg-only)
brew install ffmpeg-full
pip install whisperx

The edit, step by step

1. Gather the takes

   Put the raw clips in one folder and start Claude Code there. Tell it the goal, the length and the platform before it touches anything, and tell it to work on copies.

2. Transcribe with word timings

   Have Claude pull the audio from each clip and run WhisperX with --language set, saving JSON. Then ask it to list names and product terms it suspects were misheard.

3. Cut on the transcript

   Claude reads the words, picks the cleanest take of each line and writes the in and out points to a file you can read. Approve that list before it renders; FFmpeg then trims each segment and joins them.

4. Remove silence and filler

   Silence is mechanical: auto-editor cuts quiet stretches by loudness, and --margin keeps a little padding around each cut so the edit doesn't feel clipped. Filler words and false starts need the transcript, so that pass is Claude's. FFmpeg's silencedetect filter can confirm a gap before a cut.

5. Add captions

   Whisper and WhisperX both write SRT files. Transcribe the finished cut, not the raw takes, so the timings match, then burn the captions in with FFmpeg's subtitles filter and set the look with force_style. For word-by-word animated captions, have Claude render them in HyperFrames or with Remotion's /remotion-captions skill.

6. Layer b-roll and motion graphics

   Titles, callouts, charts and lower thirds come from HyperFrames or Remotion: Claude writes them as HTML or React, renders them and lays them over the cut. HyperFrames' registry has about 450 ready-made blocks, from transitions to charts.

7. Export for 9:16

   For horizontal footage, crop the center to 9:16 and scale to 1080x1920. A center crop can cut off a speaker who sits off-center, so check the framing or ask Claude to shift the crop shot by shot.

# 2. Transcribe with word timings
ffmpeg -i take1.mp4 -vn -ac 1 -ar 16000 take1.wav
whisperx take1.wav --model large-v2 --language en --output_format json --output_dir transcripts

# 4. First pass on silence, then confirm a gap
auto-editor take1.mp4 --margin 0.2sec
ffmpeg -i take1.mp4 -af silencedetect=noise=-35dB:d=0.4 -f null -

# 5 and 7. Burn in captions, then make the vertical version
ffmpeg -i cut.mp4 -vf "subtitles=cut.srt:force_style='Fontsize=14'" -c:a copy captioned.mp4
ffmpeg -i captioned.mp4 -vf "crop=trunc(ih*9/16/2)*2:ih,scale=1080:1920" -c:v libx264 -crf 18 -c:a copy vertical.mp4

On a Mac, plain `brew install ffmpeg` can't burn captions. FFmpeg's subtitles filter needs libass (filter docs), and Homebrew's standard formula is built without it. Install ffmpeg-full and run the caption step with $(brew --prefix ffmpeg-full)/bin/ffmpeg, since that formula is keg-only and not on your PATH. Tell Claude this up front so it doesn't rewrite the command.

frame 000/180№002

The open tools people use

You don't have to wire this pipeline yourself. These are the open projects that come up most when people search for Claude Code video editing, with what each one actually does.

If you'd rather finish in a real editor, DaVinci Resolve is the scriptable one, with a catch: Resolve 21.1 moved Python scripting to the Studio version, a $295 license. On the free version, the cleaner path is to let Claude build the rough cut and hand Resolve a timeline file, which auto-editor writes with --export resolve.

What Claude still does badly

• Footage without speech. With no transcript, Claude works from sampled frames, and that's thin. Golem.de gave Claude Cowork drone footage of a house on the Spanish coast in August 2026 and ran the test under the headline "four hours for nothing".
• Names and jargon. Whisper mishears them. The video-edit-skill's rules call out Claude coming back as cloud, and warn that Whisper can invent a word at a cut. Read every caption before you post.
• Taste. Claude can find the cleanest take. It can't tell whether a joke lands or the pacing breathes, which is why video-use proposes a plan and waits for your OK before it cuts.
• Color. video-use grades every segment with a preset look, such as warm cinematic, or a custom FFmpeg filter chain. That's a style, not shot-by-shot correction, so check skin tones and mismatched shots yourself.

Make the agent work on copies. Put "never modify or delete the originals; write everything to ./edit" in your first message. video-use keeps all its output in an edit/ folder next to your sources for the same reason.

To make a video from nothing instead of cutting one, start with can Claude make videos, and compare the two rendering frameworks in HyperFrames vs Remotion. For the same workflow in OpenAI's tools, see can ChatGPT edit videos. Formato's packs are formats you open in your agent: №002 Reelforge is out at $49, or $99 for the box set of three packs, while №001 Motion Ads and №003 Skits are on press.