AI Extensions Digest: Anatomy, the skill that makes your agent build isometric machines

What it is

Ask the agent for a figure that explains an idea, and Anatomy has it invent a physical object whose working is the concept: a cache key that skips work when nothing changed becomes a pin-tumbler lock that only turns when every pin matches; one GPU context serving many canvases becomes a gantry plotter visiting sixteen wells. The agent then builds that object from dozens of small shaded solids, painted back to front in an isometric camera, framed in a card with a title, a hint and a live readout, and set moving under the pointer and keyboard. Three output kinds share one kit: pure SVG (standalone HTML, a lone .svg, or React/Next.js components), SVG sandwiched around a WebGL shader layer for fire, water and glow, and a beta 3D mode where parts re-project and re-shade every frame as you drag. The committed examples - a Raptor engine that disassembles into eleven assemblies, a typing 1984 desk computer, a ripple tank - are built pages you can open directly.

The part that raises it above a drawing prompt is verification. A geometry audit proves nothing passes through anything, every end sits on a mount, and overlaps are drawn in depth order; headless Chrome takes screenshots and contact sheets; a lines check fails any stroke that ends in mid-air; and the skill instructs the agent to read every PNG it produces. In other words, the same loop a careful engineer applies to code - build, then attack your own work - applied to illustration.

Adoption, license, setup

The repository was created on October 8, 2026; when I checked on October 20 it had about 17,000 stars, 1,295 forks and 59 open issues - fast uptake even by agent-tooling standards. The LICENSE file is MIT, copyright "2026 Ryan". Installation is the standard Agent Skill shape: npx skills add wheresryan22/anatomy -g (the CLI asks which agents to set up, and wants Node 22.20 or later) or a plain git clone into the agent's skills folder; the whole kit is about 10 MB including the committed examples. There is nothing to npm install for the core - the drawing kit uses only Node's built-in modules - and no API keys beyond whatever agent you already run. The repo is the skill: SKILL.md holds the workflow and its rules, kit/ holds the isometric camera, solids, pipes, shaders and audits, references/ holds the long-form guides the agent opens on demand.

Limits and security notes

What it does not do: the 3D mode is explicitly beta, the densest example (the Raptor engine) takes minutes to build, and the self-checks need headless Chrome - which sandboxed agents may block. The README's Codex section is candid about this: the figure still gets built, only its checks fail, and it offers scoped fixes (approve the check command on request, or allow Chrome in the sandbox config) while warning that --sandbox danger-full-access "lifts every limit for that session, so keep it for projects you trust." That is the right instinct and worth quoting rather than paraphrasing. On privacy: the kit itself is local Node scripts with no network calls and no keys, so the data exposure is whatever your agent already sees - but as with any skill, you are handing an agent a standing instruction to write and execute build scripts and launch a browser, so install it where you would run that agent's code anyway. One governance note: the top contributor list shows the author plus a handful of small accounts, and the project is twelve days old; MIT-licensed or not, this is a single-maintainer project in its honeymoon phase, so pin a commit if you build on it.