Skills

A skill is a block of instructions. When it is installed for a task's scope, its instructions are added to the prompt of every compatible agent — so skills work with every agent, including those without MCP support.

Manifest

{
  "id": "api-conventions",
  "name": "API conventions",
  "version": "1.0.0",
  "type": "skill",
  "description": "How this company designs HTTP APIs.",
  "compatibleAgents": [],
  "permissions": ["filesystem.project.read", "filesystem.project.write"],
  "triggers": { "files": ["src/routes/**"], "dependencies": ["fastify"], "keywords": ["endpoint", "api"] },
  "skill": { "instructions": "Validate every input with zod at the route. Return errors as {error:{code,message}}. …" }
}

• compatibleAgents: empty means every agent; list agent ids to restrict it.
• triggers: hints that suggest when the skill is relevant (files, dependencies, keywords).
• recommendedMcp: MCP servers that pair well with it.

Writing good instructions

• Be specific to your stack and conventions; agents already know general best practice.
• Say what to check before finishing (which scripts to run), because verification will run them anyway.
• Keep it short enough to stay in focus — a page, not a book.
• Never put secrets in instructions.

Knowledge or skill?

Organization and project knowledge (dashboard settings) is free text for background. A skill is versioned, installable at any scope, shareable between organizations, and governed by policy.