I was a product manager who got tired of AI giving me a very confident intern's answer. So I started writing down what the senior colleague would say — one markdown file per task — and it became PM Skills: 1166 agent skills for Claude, ChatGPT, Gemini, Cursor and Codex, in Anthropic's official plugin directory, MIT, free forever.
The rule for everything I ship: it has to be measured, or it has to say it isn't. Evals in the open, negative findings published, a keyword baseline every fancy router has to beat.
 |
 |
 |
 |
…and the e-ink side quest
Firmware for the Xteink X4/X3 e-readers (ESP32-C3), because a $40 e-ink device deserves better software: Inkcards (spaced-repetition flashcards, Anki import), PocketWiki (offline Wikipedia/Kiwix reader) and Inkkit (the shared hardware layer, host-testable).
A product manager's profile should read like a product. So here is the product, documented the way I'd document anyone else's: a PRD, a roadmap, a changelog, a funnel, a status page, a retro, a teardown. Every number is live; every card rebuilds itself each morning.
|
|
|
|
- Agent skills — writing them, evaluating them, publishing them to every tool at once, and why a SKILL.md beats a prompt library
- Typed decisions — when a probability is the right output and a paragraph is the wrong one
- Evals that don't lie — blind judges, keyword baselines, and publishing the misses
- E-ink firmware — ESP32-C3, PlatformIO, and making a two-colour screen feel fast
- Building the decision layer for PM Skills: route, guard, gate and judge, with typed questions
- Keeping 1166 skills honest with a CI that counts, lints, audits and benchmarks on every commit
- Trying to make "PM stands for Professional" stick
▶ Playground · Catalogue & newsletter · Sponsor · Email
Every picture on this page rebuilds itself each morning from a GitHub Action (the profile card, the case study, the visuals; no dependencies). Steal them. The social card is the same numbers at 1200×630.
