video-talkcraft is an Agent Skill that turns Claude Code and Codex into an explainer-video studio, 766 stars in 15 days

video-talkcraft is an Agent Skill from Vincent Wei that turns Claude Code or Codex into a working explainer-video studio. It takes a voiceover script and a recorded voice track, aligns them at the character level, and renders a Remotion video whose motion, camera and sound land on the actual voice. The repository was created on 22 August 2026 and reached 766 stars and 71 forks by 6 September.

The library is 108 motion recipe cards, each with a stated intent, its parameters, its known failure modes, and a self-contained Remotion TSX source that copies out as one file. The gallery renders them all live in one page. A seven-layer camera and layout system enforces one gentle push or pull per scene plus its layering rules, so a static frame is structurally impossible; anything that slips through is caught by a static-segment detector during review.

The character-level alignment is the hard part

The script-to-voice alignment is where a talking-head video usually falls apart. video-talkcraft ships two ASR backends: FireRedASR2-CTC as the default and faster-whisper as an alternative. On a 110-second bilingual Chinese-English voiceover, the maintainer's own measurement against a GPU-based forced aligner reports median character-level drift of 20 to 40 milliseconds, worst case 200 milliseconds, and zero false alarms from the built-in quality check. Each motion beat is then anchored to a specific character in the script, so kinetic captions and animated data reveal land on the sound.

The other machinery worth naming: an anti-slideshow camera rig, six preset transition families, a face-safe zone measured from a real face bounding box detector rather than eyeballed, and a triple acceptance stage that runs a machine gate on rendered stills, a pure sound-effects track check per cue, and one independent review pass on a QA contact sheet.

The measured rendering numbers

The 2 September update introduced sectioned rendering, and the maintainer's log records the delta on a 201-second vertical test video: first full render dropped from 13 minutes to 9, changing one shot and re-emitting a sound-locked new video dropped from a full re-render to 53 seconds, 43 still frames dropped from 11 minutes to about 1, and the QA pass on 160 review frames dropped from roughly 160,000 tokens over 21 minutes to 40,000 tokens over 7 minutes by tiling into a 3 by 4 contact sheet.

The licence is the friction

video-talkcraft ships under PolyForm Noncommercial 1.0.0, which permits personal, academic and internal-evaluation use, and requires prior authorisation for any commercial deployment. A studio or agency planning to bill for videos built on this pipeline needs a licence conversation before adoption, which is a different constraint from the MIT and Apache defaults most Claude Code and Codex skills carry.

Source

Vincentwei1021/video-talkcraft on GitHub, and its live gallery of 108 motion recipe cards. Star, fork and creation dates from the GitHub REST API, verified 7 September 2026.