What it takes to make a coding agent a reliable creative collaborator

Each film in this repo is a single index.html file — no libraries, no external fonts, no network calls, no images loaded at runtime. Canvas 2D and Web Audio do everything. That constraint is the whole aesthetic: the risograph look, the procedural scoring, the deterministic frame function that makes every moment inspectable and every MP4 mathematically incapable of dropping frames.

The eight films run between 28 and 78 seconds. Window Seat is the anchor — a night train journey whose every pixel except the piano is generated, including a glass of water that leans with each acceleration and is the last thing to settle. Roost renders a starling murmuration where each bird is a 2–3 pixel ink dot, so the flock itself becomes the halftone. Eclosion builds a monarch butterfly's wings as a planar 3D model measured from an actual specimen photograph, layered by depth so the body reads correctly from multiple angles. These are not sketches. The design decisions and remaining weaknesses are documented per-work in FILM.md files.

The production method is the more interesting story for anyone building with agents. The author directed each work; Claude Code wrote the code, guided by a structured kit of skills, rules, and craft documentation living in the repo itself. The workflow is explicit: design the work in a markdown spec, prove the hardest frame first, inspect via frame strips and loudness sheets, iterate. A determinism hook warns when any edit breaks the pure-function guarantee. The skills in .claude/skills/ carry the workflow gates, and each skill's examples file points to reusable routines from the shipped films.

What this demonstrates is a particular discipline about agent-assisted creative work: the agent operates inside a tightly specified contract (CLAUDE.md), with tooling that makes regressions visible immediately. The repo isn't arguing that AI wrote art — it's showing what a human director needs to build so that a coding agent can be a reliable collaborator on work with genuine craft requirements. The quality bar is documented, the weaknesses are named, and the harness enforces the properties that matter.

The print series — Cabinet's natural-history plates built from growth geometry, Sceneries' landscapes — follow the same architecture: integer time steps, native PNG export, closed-form geometry. Everything is reproducible from source.

For anyone curious about using coding agents for generative or creative work, this repo is a rare case where the scaffolding is as carefully considered as the output. The make your own section is not aspirational — the tools, skills, and examples are all present, and the instruction to open Claude Code and ask for a lighthouse keeper's night is a genuine starting point, not a demo hook.