Anatomy is a free, open-source Claude Code skill that turns one prompt into an interactive isometric SVG machine. Here's how to install it and where it breaks.
Vamsi Tallapudi
Manager, Architect Technology at Cognizant
Anatomy is a free, open-source skill for Claude Code that turns one sentence into an interactive isometric diagram: a little machine with shaded parts, moving pieces, and live readouts that act out whatever concept you described. You install it with a single command, type /anatomy followed by what you want explained, and Claude designs the mechanism, draws the SVG geometry, and wires up the interaction code on its own.
I'd seen the screenshots circulating on X for a few days before I actually installed it, and honestly they looked too clean to be agent output. They're not. I pointed it at a rate limiter for a client doc this week and it came back with a mechanical token bucket: pegs dropping into a bin, a gate that only opens once there are enough tokens stacked up. Took about 40 minutes of Claude working on its own. That's the whole pitch of this thing: instead of boxes and arrows, it invents a physical object whose moving parts are the explanation.
Anatomy is built by a developer going by wheresryan22 and lives at github.com/wheresryan22/anatomy under the MIT license. Give it a technical concept (a rate limiter, a cache eviction policy, an escapement mechanism, a rocket engine) and it builds a real-world machine whose operation maps onto that concept, then renders it as a shaded isometric figure you can hover, drag, or click.
Three output tiers exist:
Every figure ships with a title card, a live numeric readout, and eased motion that responds to your mouse or keyboard. It's not a static picture; it's a tiny simulation.
One command, assuming you've got Node.js 20.10 or later and Chrome/Chromium installed (Chrome drives the automated screenshot checks Claude runs on itself before calling a figure done):
npx skills add wheresryan22/anatomy -g -a claude-code
That uses the skills CLI from Vercel Labs, which finds the SKILL.md at the repo root and installs it globally. Swap the -a flag for antigravity, opencode, or minimax-code if you're on a different agent. Codex uses $anatomy instead of a slash command, but the prompt syntax is otherwise identical everywhere.
Once it's installed, you just talk to it:
/anatomy explain how a token bucket rate limiter works, interactive
The skill's own docs are blunt about what makes a good prompt: name the actual concept, not "draw me a diagram," and hand it real numbers if you have them. It invents the machine; you supply the constraints.
Anatomy sizes every request into one of three buckets before it starts building, and it's upfront about the time cost: this is unattended agent work, not a chat reply.
| Size | Parts | Time | What Claude does |
|---|---|---|---|
| Figure | 40–80 | Under 1 hour | Just builds it, no heads-up needed |
| Hero | 80–150 | 2–4 hours | Tells you first, then builds a denser object with multiple systems |
| Epic | 150+ | 6+ hours | Proposes a Hero version first; won't start without your okay |
Final output lands as a standalone HTML page, a single self-contained .svg file, React/Next.js components, or a printable PDF sheet via the included sheet.mjs script. Before calling anything finished, the skill runs its own audit: 4x zoom close-ups on every joint, a line-ends check that every stroke terminates cleanly, and (for pipes and cables) a mandatory node build.mjs --audit pass. If that audit doesn't come back clean, the figure isn't done. That's a more rigorous self-check than most AI-generated visual output bothers with.
wheresryan22 posted the launch on October 10, 2026: "Describe a machine in one line. Claude brings it to life." Within days it was everywhere: midudev, GitHub's own account, and a wave of Chinese AI accounts all reposted demos of mechanical watches, a Raptor engine test stand, an arcade cabinet, a Tesla coil. One repost calling out the virality claimed 1.9M views in two days on the original post. GitHub star counts moved fast too: roughly 2,100 within three days of launch, past 16,000 by the time I'm writing this. That's a real viral curve, not an inflated one. The figures people were sharing genuinely looked like nothing else coming out of a coding agent.
The skill's own limits section is more honest than most AI tool docs, and worth reading before you lean on this for anything that needs to be exact:
Yeah, if you write docs, blog explainers, or landing pages where readers genuinely benefit from seeing a system behave like a machine. I'd reach for it again for exactly the kind of post I used it for this week — explaining a mechanism nobody intuitively gets just from a sentence.
I wouldn't reach for it for data visualization, timelines, org charts, or anything that needs to look the same twice. For those, a boring static diagram tool wins every time. And budget your Claude Code usage accordingly: a Hero-sized figure can eat several hours of agent time, which matters if you're watching token spend (see our breakdown of Claude Code usage limits and cost).
If you're trying to decide what else is worth adding to your Claude Code setup this year, it's one of the more genuinely novel entries in our running list of top AI tools for 2026. And if you're curious how Claude Code's broader feature set has been shifting lately, we covered the 2.1.221 Focus View update separately.
New AI tools, automation workflows, and course drops — straight to your inbox. Join 2,400+ builders.
HybridAgent lets LLM agents think in latent vectors and act in text, cutting tokens up to 54.7% and cost up to 3.7x. Here's what it means for agent builders.
DeepSeek's new Search and Fetch API costs $0.75 per 1,000 calls, undercutting Tavily and Exa by 10x. Here's the pricing, setup for Claude Code/Codex, and the data-residency catch.
GPT-6 Pro in ChatGPT now splits tasks across multiple agents running at once. Here's who gets it, what it actually does, and how it compares to Claude Code subagents.
Comments