After the first example, we stop and name the patterns. This is where the loose ends get tied: what bm-skill-builder is and where to get it, where skills live and when to make one global, and the agentcanon convention that lets any AI app find them. Then the anatomy of a skill itself: what SKILL.md holds, what a template does, why declared inputs matter, and why the interview pattern is the highest-leverage convention in the course. We close with the three moves you saw improve the proposal skill, and how to apply them to whatever document you keep rebuilding.
What's covered:
- The loop, run twice: Part A basic, Part B one level up
- bm-skill-builder: what it does and where to find it
- Repo-level vs global skills, and the rule of thumb for choosing
- agentcanon and AGENTS.md: one home for skills, one file that describes the repo
- The anatomy: SKILL.md, template, declared inputs, and the optional structure around them
- The interview pattern: recommend, then confirm
- Start simple, and add structure only when real work demands it
- Why shared training material lives outside any one skill
- Three improvement moves: make it ask, close the library, kill and promote