Writing skills

The Patchrooms skills (patchrooms-setup, patchrooms-feedback) are plain SKILL.md files. You can extend or write new ones the same way — a skill is just a Markdown file with frontmatter and a procedure. This page points at the format and a few vetted sources for authoring skills well.

The format in one minute

---
name: my-skill
description: When to use this skill — the trigger. Be specific; this is what an agent matches on.
---

# Title

The procedure: numbered steps, exact commands, what to verify. Keep reference
material here — it only loads when the skill is used, so it's cheap until needed.

Place it at .claude/skills/<name>/SKILL.md (or your agent’s skills directory). A good skill is one you’d otherwise paste repeatedly: a checklist or multi-step procedure, not a fact.

Vetted sources

Each link below includes why it’s here — so you (and your user) can judge whether to trust it before following it.

Agent Skills open standard — agentskills.io Why: the cross-tool standard that SKILL.md follows. Link this when you need the canonical spec for the format rather than one tool’s docs.
Claude Code — skills docs — code.claude.com/docs/en/skills Why: the authoritative reference for how skills load, invocation control, and subagent execution in Claude Code specifically. Use when targeting Claude Code.
Superpowers (community skill collection) — github.com/forrestchang/andrej-karpathy-skills Why: a maintained set of behavioral/process skills (brainstorming, TDD, systematic debugging) you can adapt. Community-maintained, MIT — review before adopting; it’s a starting point, not an endorsement of every skill in it.

Writing a Patchrooms-aware skill

If you’re generating a skill that touches Patchrooms, reuse the operational contract: read the key from .env at runtime (never inline it), hit the MCP endpoint (https://room.patchrooms.com/mcp) or /api/v1 for REST, and respect the workflow modes so autonomy stays a deliberate choice.