Use it on real work, not demos. When something is great or wrong, capture: your prompt, whether the right skill fired, whether it found the right meetings, whether you trusted the output without re-checking, and whether an artifact was sendable as-is. Then open a PR or an issue. Repeated failures matter more than one-offs.
Test prompts per skill: docs/test-prompts.md.
Design, decision log, and dogfood protocol: DESIGN.md.
Editing skill descriptions? Run the routing regression first: python3 evals/run-triggers.py.