The extension map
Claude Code's extensions fall into three groups:
- Always-on context:
CLAUDE.md, rules and output styles. - On-demand capabilities: skills, MCP servers and subagents.
- Event-driven automation: hooks.
Plugins sit on top as the packaging layer that bundles any of these for sharing.
Signs you need each one
| When you notice… | Add a… |
|---|---|
| Claude gets a convention wrong twice | Line in CLAUDE.md |
| You keep asking for a certain format or length | Output style |
| You keep typing the same starter prompt | User-invoked skill |
| You paste the same playbook for the third time | Skill |
| You keep copying data in from a browser tab or another app | MCP server |
| A side task floods your conversation with output | Subagent |
| Something must happen every time, no exceptions | Hook |
| A second repo needs the same setup | Plugin |
Head-to-head
| CLAUDE.md | Skill | Hook | Subagent | |
|---|---|---|---|---|
| Loads | Every session | When invoked or relevant | On lifecycle events | When spawned |
| Context cost | Every request | Low (description) until used | Zero unless it outputs | Separate window |
| Deterministic? | No (guidance) | No (instructions) | Yes | No |
| Best for | Facts and conventions | Procedures and expertise | Enforcement | Isolation and parallelism |
◎Pick the extensionpick the best fit
Every edit must be auto-formatted with Prettier, no exceptions.
Must happen every time, deterministically: a PostToolUse hook.
Claude should query your production metrics database.
External data access is MCP's job.
The 9-step procedure to cut a mobile release.
A long, occasional workflow, loaded on demand.
The test command is
pnpm test -- <file>.A short fact that's relevant in most sessions.
Audit 300 files for deprecated API usage and report back.
Heavy reading in an isolated context, with only findings returned.
Better together
- Skill + MCP: MCP gives Claude access to Slack; a
/post-updateskill teaches how your team writes updates. - Skill + subagents: an
/auditskill fans out to security, performance and style reviewer subagents. - Hook + skill: a
PostToolUselint hook catches problems, and a/fix-lintskill knows how to fix your linter's common complaints. - CLAUDE.md + skills: CLAUDE.md says "for releases, use
/release"; the skill holds the details.
✻
The golden rule
Start with CLAUDE.md. Add other features only when a specific need shows up. A minimal setup that you understand beats a sprawling one you don't.
Check your understanding
?Lesson quiz10 XP per correct answer · +25 for a perfect score
QUESTION 1 OF 4
Which extension is deterministic, firing without Claude deciding?
Why: Hooks run on lifecycle events every time. The others are instructions Claude chooses to follow.
QUESTION 2 OF 4
You keep pasting the same deployment checklist. What should you create?
Why: Repeated playbooks are exactly what skills are for.
QUESTION 3 OF 4
What does MCP provide that a skill doesn't?
Why: MCP supplies the access. Skills supply the know-how.
QUESTION 4 OF 4
Which has a context cost on every single request?
Why: CLAUDE.md loads in full every session, which is why it must stay lean.
0 / 4 correct