Source: Extend Claude Code (Claude Code docs)
Anthropic's map of Claude Code extension points (CLAUDE.md, rules, skills, MCP, subagents, hooks, plugins) with when-to-use guidance and per-feature context cost.
Contents
- What: official product documentation
- Publisher: Anthropic
- Date: living document, read 2026-11-14
- Original: Extend Claude Code
Summary
A decision guide: each extension plugs into a different part of the agent loop and carries a different, explicit context cost. It also gives the order in which teams typically adopt them.
Key claims
- CLAUDE.md: loaded every session, costs tokens on every request; for ‘always do X’ rules. Keep under 200 lines.
.claude/rules/: every session or only when files matchingpathsare opened.- Skills: descriptions every request, full content on use;
disable-model-invocation: trueremoves even the description. - MCP: tool names at start, full schemas deferred (tool search on by default).
- Subagents: isolated context; hooks: zero context unless they return output.
- Layering: CLAUDE.md files are additive; skills and subagents override by name; MCP servers override by name (local > project > user); hooks merge and all fire.
- Hook vs skill: a hook fires deterministically on an event; a skill is interpreted by the model. A ’never edit .env’ line in CLAUDE.md is a request; a
PreToolUsehook is enforcement. - MCP vs skill: MCP supplies connections and tools; a skill supplies knowledge of how to use them (e.g. schema and query patterns).
- Adoption triggers: a convention corrected twice → CLAUDE.md; same prompt typed repeatedly → user-invocable skill; a playbook pasted a third time → skill; data copied from a system the agent can’t see → MCP; must happen every time → hook; second repo needs the setup → plugin.