Anatomy: The Free Claude Skill That Explains an Idea by Drawing a Working Machine
Most explainer graphics lie a little. The arrow points roughly the right way, the gauge shows a number that looked nice in the mockup, and the caption says something slightly different from the picture above it. Nobody means harm. The drawing and the facts just live in two different places, and they drift.
A free tool called Anatomy is built on the opposite idea. It explains a concept by inventing a physical machine whose working is the concept, then draws that machine as a detailed isometric figure, the angled three-quarter view you know from technical manuals. The dial on the front, the parts that move, and the caption underneath all read from the same small set of numbers. Change one value and all three change together.
It went up on GitHub on October 8 and passed 8,800 stars in under a week. That is a lot of attention for something its own author describes as "slow on purpose." I think the slowness is the interesting part.
What it is, in plain terms
Anatomy is a skill: a folder of written instructions and helper scripts that an AI coding assistant reads before it starts work. It runs inside Claude Code, Codex, OpenCode and Antigravity. You do not open an Anatomy app. You ask your assistant for a figure, and the assistant follows Anatomy's rulebook to build it.
You type something like this:
/anatomy explain how a token bucket rate limiter works, interactive
What comes back is not a picture file in the usual sense. It is a small web page, plus the script that drew it, showing a shaded machine inside a framed card with a figure number, a title, a live readout and gentle motion that responds to your mouse and keyboard. You can also ask for a single vector file (the scalable format that Illustrator, Figma and browsers all open) or a component a web developer can drop into a React site.
There are two optional extras. One adds browser-drawn light effects for fire, water and glow. The other, still marked beta, makes the parts turn in true 3D.
The licence is MIT, which means you can use it for client work, change it and ship it. It needs Node.js 20.10 or later and nothing else to install.
Why "slow on purpose" is the feature
Most AI image tools compete on speed. Anatomy's instructions tell the assistant to do the opposite. The SKILL.md file, which is the rulebook the assistant reads, lays out a process that looks a lot like how a careful technical illustrator works:
- Truth first. Collect the real sizes, counts, thresholds and behaviour before drawing anything, and keep them in one small file that drives the readout, the motion and the caption.
- Invent the object. Pick a machine whose mechanism maps onto the idea. The rulebook's line for this is "The object is the explanation."
- Write a parts list. At least 40 parts, grouped into base, structure, mechanism, controls and the thing being explained, with every part resting on another part.
- Build and check. Draw the parts back to front, then run a check script until it reports "audit passed" and a second one until no line ends are left dangling. Then take close-up screenshots of every joint in both light and dark versions and read them.
That last step is the one I would point any designer at. The assistant is told to inspect its own work at four times magnification before showing it to you. The quality bar in the file is blunt: every region should still reward a 4x zoom, because "a bare face or an empty corner is unfinished work, not calm."
You can argue with any of those rules. That is the point. They are written down, so you can read them, and if your studio draws differently, you can change them. A model that improvises its style every time gives you nothing to edit except the prompt.
The size of the job is a decision you make
Anatomy sorts every request into three sizes, and it is unusually honest about what each costs.
- Figure: one mechanism, 40 to 80 parts. This is the default, and the assistant just builds it.
- Hero: 80 to 150 parts. The rulebook estimates about two to four hours of assistant work.
- Epic: six hours or more. Here the instructions say, in so many words, "Do not start it." The assistant is supposed to give you an estimate, offer a Hero version instead, and only build the Epic if you choose it.
Two to four hours of assistant time is real money on a paid AI plan, and the project publishes no figures on token use or cost. So treat that estimate as the price tag. A Figure for a blog post is a reasonable experiment. A Hero for a client deck is a budget conversation.
What you can make with it
The repo ships eight finished examples, and the built pages are already in the examples/ folder, so you can open them in a browser before installing anything. They give a good sense of range:
- a dial gauge on a stand over a carriage of gauge blocks, about 300 lines of code
- a 1984-style desk computer with a keyboard you can type on and a CRT screen
- an arcade cabinet where a dropped coin runs down the chute and the screen starts its demo loop
- a rocket engine in a test stand that comes apart into 11 assemblies, with a glowing exhaust plume
- a school ripple tank with lamp light playing through the water
If you make explainer videos, illustrate for technical publications, design onboarding screens or teach, the use is easy to picture. Think of the cutaway illustrations in an old engineering textbook. Now imagine one that moves, shows live values, and was drawn by an assistant in an afternoon from your notes.
It also helps to know what it refuses. The README says it draws "machines, instruments, tools, and furniture," not "people or exact maps," and that "It is not a chart library." If your idea is really a bar chart, Anatomy is the wrong tool.
Put this into practice
Here is how I would run a first test without burning an afternoon.
1. Look before you install. Open dial-indicator and arcade-cabinet from the repo's examples folder. Decide whether the look suits your work. It is a specific style: dark card, shaded isometric parts, small caps labels. If it does not fit your brand, stop here.
2. Install it. In Claude Code, the README gives one command:
npx skills add wheresryan22/anatomy -g -a claude-code
You can also download the folder and drop it into your assistant's skills directory.
3. Choose an idea with a real mechanism. Anatomy works best when something physically pushes, fills, turns or drains. A compressor ducking a bassline, a camera shutter, an ink cartridge running low, a queue of print jobs. Ideas that are mostly relationships or opinions will fight it.
4. Give it true numbers. The README's advice is to "name the idea, not the drawing" and supply real values. Do not describe the machine you imagine. Describe the concept and its numbers, and let the skill invent the object. For a compressor that might be a threshold, a ratio, an attack time and a release time.
5. Say where it will live. A standalone page, a single vector file, or a component for a website. Each comes out differently, so decide up front.
6. Ask for a Figure first. Not a Hero. You want to see how it interprets your idea before you spend hours of assistant time.
7. Test the numbers. The live readout is the honesty check. Move the controls and confirm the values match what you supplied. If a value is wrong, it is wrong in the one small file that drives everything, which is a much easier fix than redrawing.
8. Make the light version. Each example rebuilds in a light theme with a --light flag on its build script, which matters if your page or slide background is white.
Where it breaks
Time and money. "A figure takes a long session" is the README's own phrasing. The rocket engine build takes a few minutes of computer time on its own, and a full 3D audit runs 15 to 20 minutes for a 30-part figure. None of that includes the assistant's thinking time, which is where the cost sits.
Editing still means code. The figure arrives as source you own, a build script and the page it writes. Changing a number is easy if you are comfortable opening a text file. Moving a part or changing its colour by hand is a job for someone who reads JavaScript, or another round with the assistant.
No people, no maps, no charts. That rules out a large share of explainer work. You will still need an illustrator, a cartographer or a charting tool for those.
The effects need a modern browser. The light and water effects need WebGL support. Without it the plain figure still shows, and for viewers who turn off motion in their system settings, everything holds still.
3D is beta. A 380-part stress test ran at about 20 frames per second at 1440 pixels wide. Fine for a demo. Not yet something I would put on a client's homepage.
It is one week old. Thousands of developers have bookmarked it. Very few have used it for paid work yet, and there is no record of how it holds up on a deadline.
The rulebook is the product
The interesting thing about Anatomy is not that an AI can draw a machine. Plenty of image generators can produce a convincing cutaway. The difference is that this one comes with its working method written out: get the facts first, list the parts, build, audit, zoom in, check again. That method is something a human illustrator spends years learning, and here it sits in a text file you can read in ten minutes.
You do not have to love the house style to take something from that. If you make explainer graphics, open the SKILL.md file and read it as a description of a process. Then ask yourself which of those steps you skip when you are in a hurry, and what your own version of "audit passed" would be.
Medium metadata
Title: Anatomy: The Free Claude Skill That Explains an Idea by Drawing a Working Machine
Subtitle: A new open-source skill for Claude Code and Codex turns a concept into a detailed, moving technical illustration, and its written rulebook (true numbers first, 40 parts minimum, audit before handover) is more useful than the pictures.
Tags: Illustration, Information Design, Artificial Intelligence, Claude, Design Tools
Estimated read time: 8 minutes