CLAUDE.md vs skills vs hooks vs subagents vs MCP vs plugins
Which Claude Code extension to use: CLAUDE.md, rules, skills, hooks, subagents, MCP, and plugins compared by when they load and the trigger for adding each.
Claude Code now has at least seven ways to customize it: CLAUDE.md, rules, skills, hooks, subagents, MCP servers, and plugins. Picking the wrong one is how a team ends up with a 600-line CLAUDE.md, a hook that should have been a sentence, or a skill nobody's agent ever loads.
Anthropic's features overview explains each extension and, more usefully, the moment each one becomes worth adding. This guide condenses it, checked on 16 November 2026.
The short answer
| Extension | What it is | When it loads or runs | Use it for |
|---|---|---|---|
| CLAUDE.md | Context Claude sees every session | Every session start | Conventions and "always do X" rules |
Rules (.claude/rules/) | CLAUDE.md split by topic or path | At start, or when matching files are touched | Rules for one part of the codebase |
| Skill | Instructions, knowledge, or a workflow | When relevant, or when you run /name | Repeatable procedures and reference material |
| Hook | A script, HTTP call, MCP tool call, prompt, or subagent triggered by an event | On every matching event, regardless of Claude | Anything that must happen every time |
| Subagent | A worker with its own context, tools, and model | When Claude delegates to it | Tasks that would flood the main context |
| MCP server | A connection to an external system | Tool names at start; calls when needed | Data or actions outside the repo |
| Plugin | A package of skills, hooks, subagents, and MCP servers | When installed | Reusing a setup across repositories |
Anthropic calls skills the most flexible extension and plugins the packaging layer. New users are advised to start with CLAUDE.md and add the rest as specific needs come up. For response tone and format, there are also output styles.
Add each one when its trigger appears
The most practical part of Anthropic's guide is a list of triggers: the moment each extension earns its place. Paraphrased:
| When this happens | Add |
|---|---|
| Claude gets a convention or command wrong twice | A line in CLAUDE.md |
| You keep typing the same prompt to start a task | A skill you invoke with /name |
| You paste the same multi-step procedure into chat a third time | A skill |
| You keep copying data from a system Claude can't see | An MCP server for that system |
| A side task floods the conversation with output you won't use again | A subagent |
| You want something to happen every time without asking | A hook |
| A second repository needs the same setup | A plugin |
You don't need everything on day one. Most teams add these roughly in that order.
Hooks vs skills
The difference is who decides. A skill is instructions Claude chooses to load when a task matches its description, or that you run yourself. A hook runs on a lifecycle event, such as after every file edit, whatever Claude decides.
- "Use our API error format" is knowledge: a skill or CLAUDE.md.
- "Run ESLint after every edit" must happen every time: a hook.
- "Never edit applied migrations" must never be broken: a hook that blocks the edit.
Examples of each: Claude Code hooks: 6 examples and Claude Code skills: 6 examples.
Skills vs subagents
The difference is where the work happens. A skill adds instructions to the current conversation. A subagent works in its own context window, with its own tools and model, and returns a summary.
Use a subagent when the task reads a lot and returns a little, such as research across many files or running a test suite. Use a skill when the task is a procedure the main conversation should follow. The two combine: a skill with context: fork runs in a subagent. More in Claude Code subagents: 5 examples.
CLAUDE.md vs skills
The difference is context cost. CLAUDE.md loads in full every session, so every line is paid for on every task. A skill's body loads only when used. Short, always-relevant rules go in CLAUDE.md; long or occasional procedures go in skills. See CLAUDE.md examples for what belongs in the file.
Skills vs MCP
The difference is knowledge versus access. A skill teaches Claude how to do something. An MCP server gives Claude access to a system it can't otherwise reach: a database, Slack, a browser, your analytics. They work together. A skill can tell Claude how and when to use an MCP server's tools. For which servers to connect, see the best MCP servers for product engineers and dev teams.
Skills vs plugins
The difference is distribution. A plugin bundles skills, hooks, subagents, and MCP servers into one installable unit. Plugin skills are namespaced, like /my-plugin:review, so several plugins can coexist. Write a skill in .claude/skills/ for one repository; package it as a plugin when a second repository needs it, and share it through a marketplace, covered in Claude Code team memory. Step by step: Claude Code plugins and marketplaces.
The thing none of them is for
Every extension above holds something stable: a convention, a procedure, an automation, a connection. None of them is built for decisions that change: a feature cut from a release, a pricing rule revised in a Slack thread, a customer exception approved yesterday. Put those in CLAUDE.md or a skill and they go stale without anyone noticing, because a file has no owner per line and no record of what replaced what.
Where Lockstep fits
Lockstep handles decisions, and it is built from the same parts. It installs a Claude Code hook that briefs each session on the approved decisions that apply, and an MCP server that lets the agent query them mid-session. Decisions are captured from PRDs, Slack, and pull requests and become binding when their owner approves them; when one changes, the old version is marked replaced. For organizations, Standards & Skills rolls out versioned skills by team, repository, or path.
Session briefings and managed skills work with Claude Code today. See how it works.
Sources checked 16 November 2026
- https://code.claude.com/docs/en/features-overview
- https://code.claude.com/docs/en/skills
- https://code.claude.com/docs/en/hooks-guide
- https://code.claude.com/docs/en/sub-agents
Related reading
For how each piece loads into context, see context engineering for coding agents and Claude Code. For memory specifically, see how Claude Code memory works. Definitions: agent skills, Claude Code hooks, subagents, MCP server. For large fan-out work, see Claude Code workflows.
Frequently asked questions
- What is the difference between Claude Code hooks and skills?
- A skill is instructions Claude chooses to load when a task matches it, or that you run with a command. A hook is a script that runs automatically on a lifecycle event, such as after every file edit, regardless of what Claude decides. Use a skill for knowledge and procedures, and a hook for anything that must happen every time.
- What is the difference between skills and subagents?
- A skill adds instructions to the current conversation. A subagent is a separate worker with its own context window, tools, and model that returns only a summary. Use a subagent when a task would flood the main conversation with output. A skill can run in a subagent with context: fork.
- What is the difference between Claude Code plugins and skills?
- A plugin is packaging. It bundles skills, hooks, subagents, and MCP servers into one installable unit, with its skills namespaced like /my-plugin:review. Write a skill for one repository; package it as a plugin when a second repository needs the same setup.
- When should I use MCP instead of a skill?
- Use MCP when Claude needs data or actions from an external system, such as a database, Slack, or a browser. Use a skill when Claude needs knowledge or a procedure, such as your deploy checklist or API conventions. They combine well: a skill can tell Claude how to use an MCP server's tools.
- Should a rule go in CLAUDE.md or a skill?
- If it applies to almost every task and is short, CLAUDE.md. If it is a multi-step procedure or only matters for some tasks, a skill, because its body loads only when used. If it must never be broken, a hook.
Keep reading