A new open project on GitHub shows how an AI coding assistant can explain a technical idea by inventing a physical machine that works the same way. The project, called Anatomy, is a skill for Anthropic's Claude Code. It produces detailed, interactive isometric figures for documentation, articles and landing pages, and it is published in the repository wheresryan22/anatomy, which GitHub lists as of 11 October 2026.
What the skill does
Anatomy starts from a concept rather than a diagram. Ask Claude for a figure that explains a technical idea, and the skill has it imagine a real object whose working principle is that concept. The repository's own examples give a flavour of the approach. A cache key that skips work when nothing has changed becomes a pin-tumbler lock that only turns when every pin lines up. A single graphics processor context serving many canvases becomes a gantry plotter whose one print head visits 16 wells in turn.
Claude then builds that object from dozens of small, shaded parts. The result is placed in a framed card with a title, a short hint and a live readout. The figure responds calmly to the pointer and the keyboard, and the project says the model that moves the parts also drives every number the figure displays.
Three output formats
The figures can be delivered in several forms. The core output is SVG, with optional WebGL shaders for effects such as a glowing exhaust plume, and an experimental 3D mode. Users can ask for a standalone HTML page, a single SVG file, or a React or Next.js component, in either a light or dark theme. The project describes the light theme as the one it is designed for first.
How it is installed and used
The project is a standard Agent Skill, meaning a folder with a SKILL.md file at its root. According to the repository, it can be installed with the skills command-line tool, by cloning it into a Claude Code skills folder, or by downloading a ZIP archive and moving the unpacked folder into the user skills directory. The command-line route needs a recent version of Node.js; the documentation asks for version 22.20 or later for version 1.7.1 of the tool.
Once installed, a user types /anatomy followed by a description of what the figure should explain. Claude can also load the skill automatically when a user asks for an isometric illustration or an interactive explainer. The repository advises users to name the idea rather than the drawing, and to supply real numbers such as sizes, counts, rates and thresholds.
Built-in checks
The project includes its own verification tools. A geometry audit checks that no parts pass through one another, that every end sits on a mount, and that overlaps are drawn in the correct depth order. Headless Chrome scripts take screenshots and close-ups, and a further script flags any line that ends in mid-air. The documentation tells Claude to review every image it produces.
Limits readers should know
Anatomy is not a charting library, and it contains no axes or data series. Numbers appear only as live readouts and as the true sizes and counts of parts. It draws machines, instruments, tools and furniture, not people or precise maps, and the project notes that ideas without a clear mechanism will come out weaker.
The process is also deliberately slow. The repository says a figure can take a long session, with the Raptor engine example taking a few minutes to build and a full 3D orbit audit taking 15 to 20 minutes for 30 parts. Shaders require WebGL, and under reduced motion settings the figures stop moving. The 3D mode is described as beta, having been tested on its own example and one test figure. The repository reports that a small 3D figure runs at 60 frames per second, while a 380-part stress test managed about 20 fps at 1440 pixels wide on a heavily loaded machine.
Why it matters for developers
For technical writers and product teams, the project points to a practical use of AI coding assistants beyond generating code. Explanatory graphics are often expensive to produce, and a tool that turns a concept into a working, animated mechanism could speed up documentation and marketing pages. The repository's contribution rules, which include no comments in code and byte-identical examples after kit changes, also show a strong focus on maintainability.
The project's future will depend on how the 3D mode matures and whether the community adopts it. For now, it stands as a clear example of how a skill can extend Claude Code into visual communication, and the examples in the repository are the best place for readers to judge the results for themselves.
Related: Reduce Claude Code Token Costs With optimAIzr CLI · Anthropic's OSS Scanner Automates Security Checks for Open Source
Comments (0)