Video Editor — Communication skill for Claude Code

Video editor

A Claude skill that edits raw footage into a finished video for YouTube or phones, roughly three to twenty minutes long.

You hand it a recording where you said each line a few times. It picks the cleanest take of each, cuts tightly, keeps every cut on frames where you are facing the camera, cleans the audio, builds motion graphics in your own brand, checks its own work, and gives you the file plus captions and a publish pack. Landscape and vertical come out of the same edit.

Free and MIT-licensed. Everything runs on your machine: ffmpeg, whisper.cpp, Node and Python. No account, no upload, no per-minute cost.

Install

git clone https://github.com/gadirlabs/video-editor.git ~/.claude/skills/video-editor
cd ~/.claude/skills/video-editor/graphics && npm install

You also need `ffmpeg` (version 8 or later), `whisper.cpp` with a model, Node 20+ and Python 3.11+. `references/setup.md` has the details, and the skill checks before it starts.

Then, in Claude Code, point it at a recording:

Edit `raw/lesson-01.mov` into a course video. Brand pack is in `assets/`.

What you supply

Three things matter more than the rest, and the skill will ask for them:

**A brand pack.** Colours, fonts, logo, and a sentence describing the style. It will not invent a look for you.

**A glossary.** Every proper noun in the recording. Speech recognition mangles names every time, and without this your company name is wrong in the captions and on screen.

**How you recorded.** Do you repeat lines until you are happy? Do you read from notes beside the lens? The second switches on look-away trimming, which is the single biggest quality difference this makes.

`references/setup.md` lists everything else, including the optional [fal.ai](https://fal.ai) key for generated clips.

What it actually does

Eight phases, in `SKILL.md`. The parts worth knowing about:

**It finds the takes from the audio, not the transcript.** When you say a line four times, word time