← All resources

The guide · Claude Code video editing

Edit videos with Claude Code: the 5-tool stack.

Everyone says Claude edits their videos. Nobody shows the setup. Claude Code can't watch or hear a clip on its own. It needs tools for that: one to build the edit, one to hear every word, two to understand sound, and one to make 3D objects without paying per shot.

All five are free. The reel below was edited this way, sound effects included.

48 s · edited in Claude Code with this stack

The stack

Claude Code the agent that runs the edit. Get Claude Code (needs a paid Claude plan).
HyperFrames or Remotion. Where the project lives: the timeline, captions and motion graphics, rendered to MP4. HyperFrames on GitHub.
Whisper lets Claude hear every word and time subtitles to it. OpenAI Whisper or the quicker faster-whisper.
CLAP tells Claude what each sound is, so it can pick the right whoosh or hit. LAION CLAP.
Demucs splits a track into stems, so your voice, music and effects sit on separate layers. Demucs on GitHub.
Blender Claude builds any 3D object for free instead of burning AI credits on every shot. Download Blender, plus the Blender MCP bridge.
ffmpeg the plumbing under all of it: cuts, audio, encoding. Download ffmpeg (on a Mac, brew install ffmpeg).

The setup and the prompts

🔒 Drop your email to reveal the setup and prompts

Get the whole setup, free.

Drop your email to unlock the six setup steps, the one-shot install prompt and the starter edit prompt, plus every other guide in the library. One email a week. Unsubscribe anytime.

Set it up, step by step

Tap a timestamp to jump the reel to that tool. Prefer not to install by hand? Skip to Prompt 1 and Claude Code installs everything after itself.

01

Install Claude Code and ffmpeg

Follow the Claude Code install guide and sign in with a paid Claude plan. Then install ffmpeg, because every other tool here leans on it. Make one folder for your edits, such as ~/videos, and open Claude Code inside it.

02

Get HyperFrames or Remotion

This is where your project lives. Both build a video out of code, so Claude can lay out the timeline, captions and motion graphics, then render an MP4.

HyperFrames is plain HTML and installs as a Claude Code skill. Start here.
Remotion is React. Pick it if you already write React. Check its licence if you're a company.

You only need one.

03

Install Whisper so Claude can hear

Whisper turns your audio into text with a timestamp on every word. Claude uses those timestamps to cut dead air, find your best take and drop each subtitle on the exact word. faster-whisper gives the same result in less time on a laptop.

04

Add CLAP for sound effects

Subtitles are covered, but sound effects and music aren't. Claude can't listen to a folder of 200 whooshes and pick one. CLAP matches audio to text, so Claude can search your sound library for "short punchy whoosh" or "deep cinematic hit" and get the right file back.

05

Add Demucs to split the audio

Demucs pulls a track apart into stems: vocals, drums, bass and the rest. Claude can then clean your voice, lower the music under speech, or lift the effects out of a clip you want to reuse, without one layer bleeding into the next.

06

Connect Blender for free 3D

The hack in the video. Generating a 3D object with an AI video tool costs credits every shot. Blender is free, and the Blender MCP lets Claude Code drive it. Claude models the object once, then renders as many angles as you want for nothing.

Install Blender, then add the MCP's addon.py under Edit → Preferences → Add-ons.
In Blender's sidebar (press N), open the BlenderMCP tab and click Connect.
Register it with Claude Code. The command is in the copy block below.

07

Hand it to Claude Code and edit

Drop your raw clips in the project folder and paste Prompt 2. Claude transcribes, cuts, captions, adds sound and renders. Watch the render, tell it what to change in plain words, and render again.

The install commands

For anyone setting up by hand. Run them in your terminal, one block at a time. On Windows, use WSL.

# 1. ffmpeg (Mac with Homebrew)
brew install ffmpeg

# 2. HyperFrames as a Claude Code skill (restart Claude Code afterwards)
git clone https://github.com/heygen-com/hyperframes ~/.claude/skills/hyperframes

# 3-5. Whisper, CLAP and Demucs in one Python environment
python3 -m venv ~/video-stack
source ~/video-stack/bin/activate
pip install -U faster-whisper laion-clap demucs

# 6. Blender MCP (needs uv: https://docs.astral.sh/uv/)
claude mcp add blender uvx blender-mcp

Prompt 1: let Claude install the stack

Open Claude Code in your video folder and paste this. It checks what you already have, installs the rest, tests each tool and stops to ask before anything needs your password.

Set up this machine so you can edit my videos. Check what is already installed first and skip anything that works.

Install and test, in this order:
1. ffmpeg and ffprobe.
2. HyperFrames as a Claude Code skill: clone https://github.com/heygen-com/hyperframes into ~/.claude/skills/hyperframes.
3. A Python virtual environment at ~/video-stack with faster-whisper, laion-clap and demucs.
4. The Blender MCP (https://github.com/ahujasid/blender-mcp): check Blender is installed, register the MCP with Claude Code, and tell me exactly how to enable the add-on inside Blender.

Rules:
- Never use sudo or change system settings without asking me first.
- Test each tool after installing it: transcribe 5 seconds of audio with Whisper, run Demucs on a short clip, and score one sound file against a text label with CLAP.
- If something fails, show me the error and your fix before moving on.

When you finish, give me a table: tool, version, test passed yes/no, and anything I still need to do by hand. Then tell me to restart Claude Code.

Prompt 2: your first edit

Put your raw clips in a raw/ folder and any sound effects in sfx/. Change the bits in square brackets, then paste.

Edit the clips in ./raw into one 9:16 reel for Instagram, using HyperFrames. Target length: [45] seconds. Topic: [what the video is about].

1. HEAR IT
Transcribe every clip with faster-whisper, with word-level timestamps. If I repeated a line, keep the cleanest take. Cut dead air, ums and false starts.

2. CAPTIONS
Add word-by-word subtitles timed to the transcript. Bold, readable, inside the Reels safe zone (clear of the top 15% and bottom 25%).

3. SOUND
Use Demucs to separate my voice from any background noise or music in the clips and keep the voice clean.
Use CLAP to search ./sfx and pick a sound for each big visual moment: a whoosh on transitions, a hit when a key word lands. Keep it subtle, no more than one sound every 3 seconds.
If I give you music, duck it under my voice.

4. VISUALS
Where I name a tool, number or idea, add a simple animated card or title for it. If a shot needs a 3D object, build it in Blender through the Blender MCP instead of generating it.

5. CHECK BEFORE YOU RENDER
Show me the cut list (timestamps and the line kept) and the sound cue list (time, sound, why) and wait for my OK. Then render to ./out/reel.mp4 and check a frame at every cut.

Before you start

Film clean a quiet room and a clip-on mic save more editing than any tool here. Whisper and Demucs both do better with clear audio.
Bring your own sounds CLAP searches a library, it doesn't make sounds. Use effects you have the rights to.
Approve the cut list Claude makes mistakes. Reading its list of cuts takes a minute. Re-rendering a bad edit takes longer.
First run is slow Whisper, CLAP and Demucs each download a model the first time. After that they run offline.