Rules & skills
Teach the swarm your conventions with project rules, and reusable techniques with Agent Skills.
Two mechanisms let you shape how agents work on your project, and both live in your repo — so they're version-controlled and shared with everyone.
Rules
Rules are the conventions an agent should always follow on this project. Your repo's AGENTS.md / CLAUDE.md and every .cursor/rules/*.mdc are read verbatim into the lead agent's system prompt on every turn — not summarized, not filtered. Write down the architecture invariants, the "one right way," the things that fail review, and the agent carries them into each turn.
Edit rules in the Code tab or in Settings → Rules. They live on your repo's default branch, so a change applies to every thread.
Skills
A skill is a reusable technique — "how we write a migration," "how to debug the flaky test" — stored as a SKILL.md folder. darting.dev adopts the open Agent Skills standard.
Skills are indexed, not pasted. Every skill's name and one-line description is injected into the prompt as a cheap index; when a task matches a description, the agent reads that one SKILL.md on demand and follows it. The full body never sits in the prompt until it's needed.
Where skills come from:
| Location | Scope |
|---|---|
.claude/skills | Project — committed to your repo, shared with the team, editable in Settings → Skills. |
.cursor/skills, .agents/skills | Discovered — whatever convention your repo already uses. Read-only. |
A fresh project starts with zero skills; one appears only when you add it or when your repo already carries the convention.
They maintain themselves
A weekly Skills automation keeps the project's skills sharp — pruning stale or duplicate ones, sharpening weak triggers, and adding genuinely reusable techniques. Git history is its audit trail.
A note on trust
Discovered repo skills and anything fetched from the web are untrusted input. Only rely on skills from sources you trust, and never let a skill embed secrets or destructive commands.