Zero-asset filmmaking is tractable when agents get per-shot compute budgets

Every pixel drawn, every sound synthesised, zero media assets shipped. That is the constraint this repo builds around: a 30-second vertical film produced entirely from vanilla JavaScript canvas rendering and Web Audio API synthesis, with the output being a single self-contained HTML player and an MP4 export.

The reference film makes the ambition concrete. Seventeen shots, 32 seconds, 120 bpm, 1080×1920 at 24 fps — the monarch butterfly life cycle, rendered without a single image file or audio sample. Two visual registers alternate across the shots: hand-inked paper and navy blueprint, visible in the banner. The look is explicitly modelled on Kevin Ngo's "The life of a fruit fly," and the repo includes a reference analysis document recording exactly what was borrowed from it. That kind of documented influence is rarer than it should be.

The skill structure is worth understanding. SKILL.md describes the full production sequence. Each shot gets its own subagent, which writes a scene file exceeding a thousand lines. After that, critic waves review every shot. The README is direct about the cost: this consumes a large share of a usage plan. That is not a caveat buried in fine print — it is the second sentence under "Make a film." Anyone treating this as a cheap generative toy will be surprised.

The dependency stack is modest but specific: Node.js 20 or newer, ffmpeg on the PATH, and Chromium via Playwright for rendering. The agent requirement is real — the README names Claude Code as the reference runner, though any agent capable of dispatching parallel subagents will work. Installation is a clone and a symlink into the agent's skills folder.

Three build scripts cover the full lifecycle: check.cjs runs a six-check gate and exits cleanly when green, render.cjs writes the MP4, and build.cjs rebuilds the HTML player. The butterfly film is fully reproducible from source, which means the example is also a test fixture.

What this repo actually demonstrates is that procedural generation at the shot level — not just at the frame level — is tractable when you give an agent a well-specified scene format and enough compute budget per shot. The constraint of zero media assets is not just an aesthetic choice; it makes the output fully portable and verifiable, since there is nothing to fetch, license, or lose. The tradeoff is that every visual effect has to be expressible as canvas drawing code, which puts a ceiling on photorealism but a floor under reproducibility.