Open Skills: Portable, Composable Agent Procedures Across Every Tool

Key Takeaways

• Memory solves what your agent knows; 'procedural debt' is the second problem — the agent still doesn't know HOW you work (your voice, your testing standard, your definition of done).
• A skill is a small skill.md folder with triggers (when to use / when NOT to), the job it owns, required tools, boundaries, expected output, and a verification standard — a reusable procedure, not a one-off prompt.
• Skills are primitives; runbooks compose them into reliable workflows (e.g. voice memo → transcript → draft-in-your-voice → HTML page → published-and-verified).
• Scope matters: personal skills (your voice, your publishing habits) live globally; project skills (test commands, safe selectors) live with the repo — so the files stop drifting between Cursor rules and CLAUDE.md.
• Make verification part of the contract — 'do not call this done unless this evidence exists' — to turn agent output from review debt into real leverage.
• A session-to-skill extractor closes the flywheel: at the end of a substantial session, capture any recurring non-obvious procedure as a new skill candidate instead of losing it to chat history.