talking-head-editor
A Claude skill that edits raw talking-head footage into a finished video: tight cuts that never clip a word, studio-clean voice, reframing, word-timed captions, and cinematic motion graphics written as code.
https://github.com/user-attachments/assets/fef0180c-48fe-4862-83cf-f925b40de820
The 22-second demo above (sound on) was made by running this skill on one raw camera take, retakes and all: the cut, clean voice, grade, captions and every motion graphic. Install page: griffinwooldridge.com/skills/talking-head-editor
What it does
| Step | What happens |
|---|
| Voice | DeepFilterNet removes noise and room echo (capped so it never sounds robotic), then a gentle match-EQ and loudness to -14 LUFS. |
| Transcript | Whisper word timings, transcribed in pause-separated chunks so retakes of the same line are never merged. |
| Cut | You pick the takes. Every pause collapses to ~0.09 s, and every out point waits for the word's real end, including the "-ence" or "-ts" that comes back after a tiny dip. The cut is audited for clipped tails and leftover silence. |
| Framing | Face found automatically. Full shots at 110 %, optional punch-ins, and a split layout (face left, graphic right) under panel graphics. One lanczos pass from the camera original, with an optional LUT. |
| Motion graphics | Each graphic is a small JS file whose render(t) is a pure function of time, timed to words with wt("word"). Rendered with HyperFrames at 4-8x the frame rate and averaged into real motion blur. |
| Captions | Word-timed cards with the spoken word highlighted, under the face during split shots, hidden during full-frame graphics. |
| Compose | Frame-exact assembly, frame count and decode verified. |
The motion standard the skill holds itself to is in references/motion.md: object chains instead of scene cuts, the mechanism shown literally, a camera that moves only on beats, a cursor as the actor, and real footage inside the graphics. references/techniques.md adds the build techniques learned from studying prompt-motion.com: summed closed-form springs, two-edge springs, flood/iris/goo/liquid-glass transitions, velocity blur, a per-run brief with a banned list, and a QA scan for single-frame pops. Every graphic is restyled to the topic's brand first (references/look.md, scripts/extract_brand.js).
Install
Requires macOS or Linux, ffmpeg, Node 18+ and Python 3.9-3.12.
git clone https://github.com/griffinwooldridge/talking-head-editor ~/.claude/skills/talking-head-editor
~/.claude/skills/talking-head-editor/scripts/setup.sh
If your default python3 is 3.13 or newer, point setup at an older one: PYTHON=/path/to/python3.11 ./scripts/setup.sh.
Use
In Claude Code, from a folder with your raw recording:
/talking-head-editor edit source.mp4
Claude follows SKILL.md: clean the voice, transcribe, choose takes, cut, plan the graphics, reframe, write and critique each graphic, render, caption and compose into out/final.mp4. Give notes the way you would to an editor ("tighten the intro", "that graphic is too busy", "rethink the outro") and it edits the scripts and re-renders.
Layout
SKILL.md the workflow Claude follows
scripts/ voice, transcribe, cut, render_cut, captions, compose, qa, extract_brand, setup
mg/kit/ cine.js + motion-kit.js (seek-safe motion helpers), cine.css (the look)
mg/examples/ three full graphics to learn from
references/ motion standard, techniques, run brief, cutting rules, restyling for a brand
License
MIT. Fonts (Fraunces, Geist) are under the SIL Open Font License, see mg/fonts/OFL.txt. GSAP is downloaded at setup under its own license.
