About
or: why does this exist and who hurt you
Agents forget everything between sessions: you know this.
You've re-explained your own codebase to the same model four times this week; you've watched it make the same mistake it made yesterday, with the same confidence, in the same file.
Skillipedia is the encyclopedia we built because we got tired of being the memory.
Every entry here is something an agent needed to know, forgot, and wasted time re-learning.
So we wrote it down in a format agents can actually consume. Because apparently that's our job now.
Entries come from brunnr, a gated skill registry for agent tools. Each one passes a 26-rule security scan before it gets listed here.
We're...paranoid like that.
Taxonomy
Three types of entry. Nobody gets in for being theoretically interesting.
Skills
A repeatable capability an agent can execute. Concrete, testable, installable. If you can put it in a SKILL.md and an agent can run it without asking what you meant, it's a skill.
e.g.
ax-rubric scores tool descriptions 0-5 on whether agents will actually call them. It doesn't suggest. It scores, explains, and rewrites.
Patterns
A structural solution that shows up across codebases. Not a library. Not a framework. A shape that works, extracted from code that survived production.
e.g.
Wrap external API calls in error boundaries with backoff. Not because a blog told you to. Because 3am pages taught you to.
Learnings
Something we got wrong, then got right, then wrote down so we'd stop getting it wrong again. Learnings have a confidence score because some of them might still be wrong. We're honest like that.
e.g.
Test fidelity matters more than coverage percentage. We learned this the fun way: 100% coverage, zero caught bugs.
Domains
Nobody sat down and designed a taxonomy. Entries show up; if enough of them cluster around the same problem space, that's a domain. We don't name them until they name themselves.
Browse them on the domains page.
Confidence scores
Every entry has a confidence score.
It's not how sure we are that it's true: it's how many times it's survived contact with production without someone yelling at us.
95% means battle-tested. 75% means it works but we're watching it. Below 50% means we wrote it down before we forgot, and we're not totally sure we were right in the first place.