Tech Teardown 023｜anatomy: A 5-day 6300-star agent skill pack, turning "explaining clearly" into an interactive machine

October 8th, a repository was created, and in just five days, it earned over 6,300 stars and 450 forks. This is the achievement of wheresryan22/anatomy on GitHub. It isn’t an application or library, but a skill package: a 173-line SKILL.md document, a set of isometric rendering tools, and eight complete templates. With npx skills add, it can be installed into any agent like Claude Code, Codex, OpenCode, or Antigravity, allowing the agent to create interactive isometric illustrations that explain concepts clearly, using pure SVG as the base, with optional WebGL shaders and 3D rotation. The official description literally states, “An agent skill for creating detailed, interactive isometric illustrations of real-world machines and complex systems.” Yesterday, wordcraft was disassembled to show the output produced by the agent, while anatomy sold the methodology behind that output. By reading it over two days, you’ll fully understand the two layers of the agent ecosystem. This article breaks down the content into six sections, following the old format.

I. What is it?

The complete form of anatomy consists of three components: “specification + toolbox + templates.” The specification is SKILL.md, a workflow document written for the agent, containing everything from design philosophy to hierarchical pricing and acceptance checklist. The toolbox is the kit/ directory, with over 15,000 lines of JavaScript, each dedicated to isometric projection, rotating bodies, pipeline wiring, WebGL shaders, 3D modes, and geometric auditing. The templates are found in examples/, consisting of eight complete examples, from a 300-line dial gauge to a raptor rocket engine test bench with WebGL flame visuals. Each can be rebuilt with a single command. The entire repository has 147 files and 29,078 lines (local statistics, including examples), under a MIT license. The README badge claims no dependencies, and construction uses only Node’s built-in modules, which matches the actual package files.

What it depicts has a uniform appearance: a delicate machine from an isometric perspective, with four-color shading, painterly algorithm sorting, and card-style frames (with numbers, titles, interactive hints, and real-time readings). Movement is controlled with gentle animations. The eight templates cover varying levels of difficulty: desk-computer is a desktop computer capable of actual typing, with typed text appearing on a CRT screen inside the case, featuring about 85 entities and 850 paths; arcade-cabinet is a retro game cabinet section with removed side panels, consisting of about 240 entities and 15 cables. A coin is dropped into the slot, and it slides down the money channel at a real speed on a 30-degree inclined surface, hitting microswitches, and electrical pulses travel along the cables to the motherboard. The CRT preheats, performs self-checks, and enters an attraction mode, all simulated physically; raptor-engine is a raptor engine mounted on a test stand, with WebGL rendering of flames, shock waves, spray, and steam, sandwiched between two layers of SVG.

Its popularity path is completely different from wordcraft: wordcraft became popular through discussion posts with over three hundred stars on HN, while anatomy only had one submission with two votes of support on HN. Its dissemination follows GitHub trending, X, and skill websites (skills.wheresryan.sh/anatomy with a community gallery). Two completely different paths under the same popularity metric.

II. Repository Structure

The “soul” is in SKILL.md, line 173, where YAML frontmatter declares compatibility (written for Claude Code and also works in Antigravity, OpenCode, and Codex; requires Node.js 20.10+ and Chrome/Chromium for testing). The main text provides the complete workflow for the agent: ten steps, from scale determination, collecting real numbers, inventing objects, writing a parts list, to constructing geometry, sorting and drawing, installing card frames, adding animations, and reviewing like a critic.

The “muscle” is in kit/, line 15870. iso-kit is for equidistant projection core, lathe.mjs generates rotating bodies in any axis (bell mouths, flanges, valve bodies), tube.mjs cuts three-dimensional pipes and cables into segments that can be deeply sorted, gl.mjs is the WebGL layer, turn series contains three files for 3D rotation mode, audit.mjs performs geometric auditing, and under react/ there are server-side security drawing components and client-side animation hooks. references/ contains nine layered documents with 2247 lines each: craft, kit, webgl, 3d, motion, react, verify, walkthrough, publish, with webgl.md being the longest at 473 lines.

The peripheral parts are not lacking: evals/evals.json is the skill evaluation set, tests/fixtures/ contains four PNG visual regression baselines, .github/ contains CI and PR templates, agents/openai.yaml is Codex configuration. The evaluation set deserves a closer look: five sets of “prompt to expected output” questions, from token bucket limiters, database connection pools to CDN edge caching, gas flame, espresso machine sections, with the expected outputs filled with acceptance criteria, such as “all curved parts should be smoothly shaded without stepped color blocks” and “node build.mjs —audit must end with audit passed, with zero counts for each failure.” Providing an evaluation set for the agent skills is rare in skill packages.

III. Core Mechanism

Three things deserve mention.

The first point is the design philosophy of “the object is the explanation”. In SKILL.md, it is boldly stated: The one idea that matters: The object is the explanation. No diagrams, arrows, or annotations are allowed. Instead, create a real machine whose operation is the concept itself. The document provides a mapping from concepts to machines: a GPU shares context services across multiple drawboards, functioning as a gantry drawing machine; a print head visits 16 holes in turn; if the cache key is hit, it skips redrawing, resembling a dice lock where each pin must align before the lock rotates. The 49×49 separable blur is simplified to 49+49, forming a nailbed seal with two comb tracks. Even failure modes are outlined: too abstract (stacked boards and floating squares cannot convey any meaning) or too dull (a large board drawing several lines without volume or parts). The solution is the same: create a recognizable machine composed of many small, precise parts.

The second point is —audit frame-by-frame geometric proofing. This is the most challenging engineering highlight in the entire project. kit/audit.mjs requires recording the actual shape of each part during construction, then performing machine proofs for each frame in static, exploded, and animated states: no parts passing through other parts, every endpoint landing on the correct mounting positions, no bending or flattening, all overlaps drawn in real depth order. If problems are found, the process terminates immediately. The exploded animation of the Raptor engine is broken down into 11 assemblies with frame-by-frame depth reordering, all of which can be audited. SKILL.md also blocks cheating paths: using a fake shape to deceive the audit itself is a bug, and the audit will mark it as failed. There’s also a corresponding honesty clause: every number shown in the diagram comes from the same model that drives the parts’ movement. Readings are not fixed; they are calculated in real-time by the physical model. The speed of coins sliding down the arcade cabinet, the frequency changes in the water wave grooves—all these are based on real models.

The third point is to include work hours in skill documents. The first step of the workflow is setting a scale: for the Figure level, create a mechanism with 40 to 80 parts, typically completed within an hour. The current rule is to not provide quotes to users (this was changed in PR #43 merged on October 13); for the Hero level, create 80 to 150 parts, taking about 2 to 4 hours, and inform users before starting; for the Epic level, start within 6 hours. The rule is not to start directly; first provide an estimate and propose a Hero version that retains core creativity. Only after user selection will the full version be created. Regardless of the level, there must be a page available in the browser within the first hour, then gradually deepen the process. Having the agent report work hours to users before starting is a rare example of project management discipline written into skill standards.

IV. Quality of Engineering

Several house rules are written in CONTRIBUTING; they are strict. No comments in the code: no file header comments, no difficult explanations next to lines of code, no JSDocs. Explanations should be placed within the names and structure; only tool instructions like “use client” are exceptions. Changes to the kit must be opt-in: new options or functions’ default behaviors must equal the old ones, and all existing examples must remain byte-identical after reconstruction (this must be checked by CI). If you want to make changes, explain them in a PR and submit the rebuilt pages along with it. Byte-identical is a hard regression line that completely decouples “toolbox changes” from “workplace modifications,” ensuring that agents who modify the kit extensively won’t secretly destroy any published figures.

The rest of the infrastructure is also well-organized: four PNG visuals return to the baseline, locking in rendering of entities and dot arrays under both light and dark themes. The CI workflow is hosted on .github, and the evaluation set includes five sets of questions with acceptance criteria. Contribution structure (snapshots 10-13): 8 contributors have made a total of 119 commits, with main contributor wheresryan22 accounting for 72 or 61%, and second contributor rudycelekli with 33% or 28%. PR numbers have reached #43 over five days. It was personally initiated, but the community has already joined in; this is a different structure from wordcraft’s “one person drives three days of fifty thousand lines” model repository. The repository also creates a complete loop for community contributions: the anatomy-community repository serves as a gallery for works. When users say “publish this figure to the Anatomy community gallery,” the agent will ask for the title and GitHub name, and after confirmation, a PR will be automatically opened without any need to touch git. The skill pack even handles distribution and community contributions for users.

Five, Boundaries and Costs

The Limits section of the README is very clear about its intentions. It is not a chart library: no axes, no sequences, no data plotting. Numbers only appear as model-driven real-time readings and true sizes of components. It only depicts machines, instruments, tools, and furniture, not people or precise maps. An idea without mechanisms to build it will be weak. “Slow is intentional” (It is slow on purpose is the original text of the README): a single image requiring completion of the component list, geometry, audit, and close-ups of each joint under both themes is a long conversation; Raptor Engine construction takes several minutes, and a complete 3D audit of 30 components takes 15 to 20 minutes.

The 3D mode is marked as beta: The engine has only been tested on its own examples and a test image, and has not been deployed in production images yet. The stress testing model for 380 components (a robotic hand on a messy workbench) runs at approximately 20fps at 1440px, and only 13 to 15fps in the mobile limit settings (390px, 4 times CPU throttling), both according to the official reported test criteria. The WebGL effect depends on browser support, but fortunately SVG remains fully readable without WebGL. All images automatically stop moving under reduced motion, ensuring accessibility. There is also an implicit cost for users: this set of skills targets agents rather than humans. After reading 173 lines of guidelines and nine references, you then write geometric shapes yourself—this is not how it was designed to be used.

VI. Judgment

What’s most worth paying attention to in Anatomy is not the illustration itself, but its role as a distribution unit. If a skill package quickly rises to the top of GitHub trending, it indicates that the infrastructure layer of the agent ecosystem is beginning to produce its own star projects: previously, open-source applications and libraries were what drove rankings, with users being the focus. Now, what drives rankings can be “the methodology for teaching agents how to work,” with users being the agents themselves and humans responsible only for verification. This is exactly the opposite of wordcraft, which demonstrated that agents can produce compliant application code in large quantities. Anatomy proves that the methodology for producing quality code can be packaged, distributed, evaluated, and iterated upon. Proofs are also present: Anthropic has made Agent Skills in an open format, with Codex, OpenCode, and Antigravity already supporting loading. A skill package can be written once and run four times, becoming more similar to the situation of npm packages of the past.

It’s suitable for three types of people. Those who write agent skills: The structure of SKILL.md can be directly copied, including design philosophy, failure patterns, hierarchical pricing, workflow steps, and acceptance checklist—everything is available. evals.json even provides a model for “how to test skills.” Those involved in technical communication: “Object as Interpretation” is a complete content methodology that maps abstract concepts into recognizable machine code. This approach holds true even without its code. Those using Claude Code or Codex: With this tool, agents can generate Linear-style interactive illustrations for blogs, documents, and landing pages, but the cost is waiting for a long conversation.

Risk considerations: With a single-person-led community that is just starting out, the maintenance pressure of a 15,000-line kit will continue to increase. The frame rate of the 3D mode does not meet standards, and the beta label cannot be removed in the short term. The effectiveness of the skills highly depends on the execution ability of the host agent; switching to a weaker model may significantly degrade results. Although the evaluation set exists, public performance scores have not been released yet. But even if the repository stops updating tomorrow, it has already answered one question: Open-source star projects in the agent era can be just an excellently written guideline. 173 lines of SKILL.md, along with the honest clause that “every number comes from the same model,” make it more like an engineering work than many complete application documents.

Information Source

• Repository and Documentation: README, SKILL.md, CONTRIBUTING.md, evals/evals.json, kit/audit.mjs from github.com/wheresryan22/anatomy (local shallow clone for file count and line count using find+wc)
• Stars, forks, contributors, and commits: GitHub repos/contributors/commits API, snapshot on 2026-10-13
• HN dissemination: Algolia search, only 1▲ submission, popularity tracked based on GitHub trending and skill sites
• Hierarchical pricing, frame rate, build time: Self-reported metrics in SKILL.md and README, text marked throughout