// the find
mitsuhiko/agent-stuff
These are commands I use with agents, mostly Claude
Armin Ronacher's personal config package for the Pi Coding Agent — a collection of skills, extensions, prompt commands, and themes he uses daily. It's published to npm as `mitsupi` but is explicitly tuned to his setup: Austrian transit APIs, his Google Workspace credentials, his GitHub orgs. If you use Pi and share enough of his workflow overlap, pieces of this are worth stealing.
The skill abstraction is clean — each skill is a SKILL.md describing the interface plus shell/JS helpers, which means an agent can discover capabilities without reading code. The uv extension is particularly well thought out: it intercepts pip/python/poetry shims to steer agents away from non-uv workflows, solving a real problem where agents default to whatever Python tooling they learned in training. The web-browser skill driving Chrome via CDP rather than Playwright is a pragmatic choice — no browser binary management, works against whatever Chrome is already running. The session-breakdown extension giving actual token/cost analysis over historical sessions is something most teams build internally and never ship.
This is one person's dotfiles with a GitHub star count, not a reusable library — the README says as much, but the npm publish and 2.6k stars will lure people into thinking otherwise. There's no abstraction layer between 'Armin's credentials/paths' and 'the skill logic', so forking any Google Workspace or transit skill requires surgery, not configuration. The whole thing is load-bearing on Pi, a specific agent platform with its own SDK and extension model — none of the extensions work outside Pi, and the skills only work if the agent knows to load them. Zero tests anywhere in the tree.