Creative Coding

A Claude Code Skill Ships the Working Knowledge of Generative Art, Not the Code

Generative artist Camille Roux's genart-skill plugin is pure documentation — seven reference sheets on determinism, resolution-independence, and platform conventions that teach an AI coding agent how on-chain generative art actually works.

Most tools that promise to help an AI write generative art ship a library: starter code, boilerplate, a wrapper around p5.js or three.js. genart-skill, a free plugin for Claude Code built by generative artist Camille Roux, does the opposite. Its README states the design choice plainly: “The plugin ships no code — only Markdown.”

Knowledge instead of boilerplate

Instead of code, genart-skill is a set of seven reference sheets that teach an AI coding agent the working knowledge of the field — the parts that are easy to get subtly wrong if you’ve only read a few tutorials. Roux’s own list: determinism (seeding a PRNG from a token hash so the same input always produces the same output), resolution (rendering the identical composition correctly at any output size), features (designing traits and rarity tables the way long-form generative collections actually use them), plus sheets on ethics, tooling, verification, and platforms.

The reference sheets load on demand rather than all at once, and the platform sheets are deliberately thin on specifics — no hard-coded version numbers or field names, just the mental model, canonical documentation links, and the right questions to ask those docs. The README is explicit about why: “Claude fetches the real documentation before writing platform code,” rather than trusting a snapshot that will inevitably drift out of date.

It checks its own determinism claims — carefully

One of the more interesting design choices is how the skill handles verification: it has Claude write a small Playwright test harness directly inside the user’s own project, following its own verification.md reference sheet, to check that a sketch actually reproduces the same output on repeat runs. Notably, the README doesn’t overclaim what that check proves: “Nothing here can prove cross-machine determinism, and this plugin does not claim to.” What it verifies is same-machine reproducibility, perceptual stability across different canvas sizes, and feature stability — a narrower, honestly-stated guarantee rather than a marketing one.

Built for the platforms that actually enforce these rules

The platform sheets name the places where getting this wrong has real consequences: Art Blocks, ABX, 256ART, Verse, Highlight, Plottables, bootloader.art, plus self-hosted work and screen-based (non-minted) output. Long-form generative art platforms like these mint a piece from code and a seed, so a sketch that isn’t actually deterministic, or that breaks at an unexpected resolution, isn’t just a bug — it’s a piece that doesn’t render as sold. That’s the exact failure mode this skill’s reference sheets are built to prevent.

Roux is a working generative artist and the maintainer of awesome-generative-art, a widely used curated list of generative art platforms, libraries, and resources — background that shows in how specific and platform-literate the reference material is, rather than reading like generic AI-tool documentation.