One line in Claude Code built a working watch movement. You can wind it, slow it down, and take it apart. The skill behind it is called Anatomy, and it is the most unusual diagram tool I have seen for Claude.
What it is
Anatomy is a free, open source Claude Code skill by wheresryan22. You name an idea, and instead of boxes and arrows, Claude invents a physical machine whose moving parts are the idea. The README puts it simply: "A Claude skill that explains an idea by building it as a real, working object."
The repo was created on October 8, 2026, and v1.1.0 shipped on October 12. When I checked on October 14 it had 9,471 stars and 677 forks. By October 19 the GitHub API showed 16,774 stars. It is MIT licensed and has zero dependencies.
The idea becomes the object
The examples on the site show how it thinks:
- Bloom filter: a fruit machine. The three reels are the three hash functions, and it only pays out when all three flags are up.
- LRU cache: a bookcase.
- Cache key that skips work when nothing changed: a pin-tumbler lock that only turns when every pin matches (from the README).
Each figure is built from dozens of small, shaded parts, sits in a framed card with a title and a live readout, and moves under your pointer and keyboard. Output is SVG, SVG plus WebGL for things like fire and water, or a 3D mode that is still in beta.
The part I like most: the numbers are real. Every readout in a figure comes from the same small model that moves its parts, so the numbers are not decoration. The Raptor 2 example comes apart into 11 assemblies when you hover over it and fires with a WebGL exhaust plume. The README says an audit proves that no pipe passes through anything, at rest, apart, and in every frame between.
It is slow on purpose
These figures take a long time. The author publishes Claude's active time from the first prompt to the shipped figure (review rounds included, idle gaps left out):
| Figure | Claude's active time |
|---|---|
| Watch movement | 1 h 54 min |
| Seasons | 1 h 57 min |
| Raptor engine | 10 h 33 min |
| Turbofan | 14 h |
These are the author's own numbers, and I have not timed a build myself. The site also notes that a first working version comes quickly (38 minutes for the Raptor). Most of the time goes into the close review of every joint.
How to install it
One command, then one prompt:
npx skills add wheresryan22/anatomy -g -a claude-code
/anatomy explain how a token bucket rate limiter works, interactive
Swap -a claude-code for -a codex or -a opencode to use it in Codex or OpenCode (in Codex you type $anatomy). You need Node.js 20.10 or later, and Chrome or Chromium for the self checks. The skills CLI itself may ask for a newer Node version. The skill is free, but the agent it runs in (Claude Code, Codex) is not.
The catch
Read the Limits section before you start. It is not a chart library: no axes, no data plots. It draws machines, instruments, tools and furniture, not people or exact maps, and an idea with no mechanism to build comes out weaker. A long build also costs a long agent session, which is real usage on your plan.
My take
For docs, blog posts and landing pages that need to explain how something works, this is a genuinely new kind of output. I would not use it for a quick diagram. I would use it when one great figure is worth an hour or more of Claude's time. If you want a skill that makes Claude's regular UI work look less generic, see my post on Hallmark.
Sources: Anatomy on GitHub and the Anatomy site and collection