A Claude Code skill that renders sourced explainer videos in code, no generative AI needed

Give this tool a topic and it produces a narrated motion-graphics explainer video — 1280×720 H.264, synchronized subtitles, chapter cards, a progress bar — without touching a single stock clip or generative video model. Every frame is a pure function of its frame number, written in React and TypeScript via Remotion 4. The visual language is deliberate and fixed: black canvas, white line art, purple accents, ultra-bold headlines, with a single configuration switch between a star-field and a dot-field backdrop.

What makes this interesting for agent builders is the architecture. This is not a CLI you invoke; it is a skill folder that Claude Code or Codex reads and executes. The agent does the research (with source URLs), writes the narration, runs TTS to get per-word timestamps, storyboards every shot, then dispatches parallel build agents — one Remotion component per shot group. QC agents review rendered frames against written criteria before delivery. You are consulted at exactly four checkpoints: length and language, narration sign-off, voiceover choice, and a 30-second preview after the first build group. Everything else runs unattended.

The timing table is honest about cost. A 3-to-5 minute reference-tier film — the sweet spot, roughly 40 to 50 shots — takes about 2 hours of wall clock with 8 parallel build agents and around 2 GB of disk. Longer films (5 to 8 minutes, 60 to 80 shots) need 10 to 14 agents and up to 3 GB. The repo recommends keeping at least 5 GB free because several agents bundle Remotion simultaneously.

The reference film shipped with the repo covers RAG and knowledge bases: an English cut at 5 minutes 2 seconds (785 words, voiced by kokoro-82m running locally on CPU) and a Chinese cut at 4 minutes 54 seconds (1,490 characters, Volcengine TTS with forced alignment). Both share one storyboard and 44 shots; the English version re-times each shot to the English voiceover. The full paper trail — research document, narration, storyboard, per-shot source, QC reports, delivery notes — lives in the examples directory.

The originality constraints are worth noting. Every fact shown on screen must trace to a source URL in the research document. Every frame is drawn from scratch in code; the README explicitly acknowledges that the visual vocabulary was learned from a specific Douyin creator and states that none of their assets were used. The noncommercial license covers the toolkit; the videos you produce are yours.

The hard limits are real. Only Chinese and English are supported. The template is landscape-only. Once narration is voiced, the words are frozen — frame numbers are hard-coded into shot components, so any rewrite re-times everything downstream. And this is macOS-first: the shell scripts are zsh plus Python 3, verified on macOS, with Linux listed as probably fine and Windows untested.