talking-head-reel

A Claude Code skill that turns a messy phone recording of you talking to camera (every line said three times, false starts, pauses, "okay, again") into a vertical reel for Instagram, TikTok or Shorts: the best take of each sentence, snap zooms on the cuts, word-timed captions, logos and callouts that land on the words you say, a few memes, an end card. One prompt, one recording, an .mp4 on your Desktop.

What you need

• macOS with ffmpeg (brew install ffmpeg)
• the openai-whisper CLI (pip install openai-whisper; the turbo model downloads about 1.5 GB the first time)
• Node 20 or newer
• Claude Code

Install

git clone https://github.com/mariagorskikh/talking-head-reel ~/.claude/skills/talking-head-reel
cd ~/.claude/skills/talking-head-reel/remotion && npm install

Then, in Claude Code, hand it your recording:

Use talking-head-reel to edit ~/Downloads/IMG_1234.MOV for Instagram. I did a lot of retakes, pick the best parts, remove the pauses, add animations and zoom in and out, make it entertaining.

Paste the script you meant to say if you have one; it helps the take selection. Expect 10 to 15 minutes of Claude working, most of it the render.

How it works

1. scripts/prep.sh transcodes the portrait MOV to 1080x1920 at 30 fps so Remotion can seek it cheaply.
2. scripts/transcribe.py runs whisper with word timestamps on the whole recording and prints the table of everything you said, every take.
3. Claude picks the best take of each sentence (references/take-selection.md) into reel-segments.json; scripts/cut.py turns that into captions in edit time and a timeline.
4. Claude plans the beats (spoken word to overlay) and writes a Remotion timeline like assets/Reel.example.tsx with the components in assets/reel-overlays.tsx: chip rows, big callouts, a list that gets stamped, a quote card, a typed prompt, polaroids of your other takes, a big logo or emoji next to your head, memes, an end card.
5. scripts/stills.sh renders check frames at the key moments; scripts/render.sh renders the reel.

What is in the box

• SKILL.md: the instructions Claude follows.
• scripts/: prep, transcribe, cut, stills, render, contact sheet.
• assets/: an example timeline and cut (placeholder timings of a fictional recording, so they compile), the overlay components.
• references/: how takes get chosen, the vertical layout, Instagram safe zones, the zoom grammar.
• remotion/: a minimal Remotion 4 project with the components, the Geist fonts (SIL OFL 1.1), a small sound kit generated by remotion/scripts/make_sfx.py, and logos from simple-icons (CC0).

Notes

• The example timeline expects a recording at remotion/public/talk/ig1080.mp4 and memes in remotion/public/memes/; neither is in the repo. Run the pipeline on your own recording.
• Memes are fetched at edit time from the imgflip template API and are not redistributed here.
• Paths in SKILL.md that say src/, public/ or out/ are inside remotion/.

Built with Claude Code. MIT.