🎨 Neural network drew a machine that shows how it forgets
The 'Faktosborka' channel showed how Claude, using the Anatomy skill (MIT), built an interactive diagram of a context window: a bench with books and a glass frame — inside is what the model sees, while old 'books' fall out of the frame as the mouse moves. The agent assembled the drawing in about an hour.
🌍 Agent skills (SKILL.md format) turn LLMs into generators of verifiable code rather than images: SVG geometry, WebGL shaders, and interactivity. Concepts can be explained with live, editable artifacts rather than static diagrams.
👤 It is installed with a single command (npx skills add wheresryan22/anatomy -g) in Claude Code, OpenCode, Codex, or Antigravity and draws concepts — from rate limiters to CDN caches. Limitations: only machines and tools, not people or maps; slow by design.
Source 1: https://github.com/wheresryan22/anatomy Source 2: https://cellcog.ai/blog/anatomy-claude-skill/
