The short answer
| What it is | When it loads | Use it for | |
|---|---|---|---|
| Skill | A folder with a SKILL.md: instructions, plus optional scripts and reference files |
Its name and description are always listed; the body loads only when Claude decides the task matches, or when you type /skill-name |
A repeatable way of doing a kind of task |
| CLAUDE.md | A Markdown file of project facts and rules | Every session, in full | Things Claude must always know about this project |
| Slash command | A Markdown file in .claude/commands/ |
When you type /name |
The older format of a skill; it still works, but skills can also hold supporting files |
| Subagent | A separate agent with its own prompt and its own context window | When Claude or you hand it a task | Work that should run in isolation and report back |
| Plugin | A package that can bundle skills, subagents, hooks and MCP servers | When the plugin is enabled | Sharing a whole setup, such as Superpowers, in one install |
| MCP server | A program that gives Claude new tools, such as reading a database or calling an API | Its tools are available while it is connected | Giving Claude access to a system it cannot reach otherwise |
A skill teaches Claude how to do something. An MCP server gives Claude access to something. A plugin is a box that can carry both. A subagent is another worker that can use them.
Skills vs CLAUDE.md
Both are plain Markdown instructions. The difference is cost and timing. CLAUDE.md is loaded into every session, so every line costs context whether it matters for the task or not. A skill costs almost nothing until it is used: Claude sees only its name and description, and reads the full body when a task matches.
Anthropic's documentation puts the rule simply: create a skill when a section of CLAUDE.md has grown into a procedure rather than a fact.
Skills vs slash commands
In Claude Code a Markdown file in .claude/commands/ is the older way to make a /command. It still works and accepts most of the same frontmatter. A skill does the same job and more: it lives in a folder, so it can carry scripts and reference files, and Claude can invoke it on its own when the description matches. The folder name becomes the command, so moving .claude/commands/deploy.md to .claude/skills/deploy/SKILL.md keeps /deploy working.
To keep a skill from running unless you type its name, add disable-model-invocation: true to its frontmatter. Matt Pocock's grill-me works this way.
Skills vs subagents
A subagent is a separate Claude with its own instructions and a fresh context window. It is useful when a task would flood the main conversation, such as searching a large codebase, and you only want the result back.
The two combine in both directions:
- A skill can run in a subagent: add
context: forkto its frontmatter and Claude Code starts a subagent with the skill as its prompt. That subagent does not see your conversation, so the skill must stand on its own. - A subagent can use skills: skills can be preloaded into a subagent so it starts with that knowledge.
Superpowers' subagent-driven-development skill is an example of a skill that orchestrates subagents. See the Superpowers collection.
Skills vs plugins
A plugin is a distribution package. It can contain skills, subagents, hooks and MCP servers, and installs them together. Skills inside a plugin are namespaced, so a review skill in a plugin called my-plugin runs as /my-plugin:review.
Many popular skill collections ship both ways. Superpowers, Matt Pocock's skills and Anthropic's skills can each be installed as a Claude Code plugin, which updates itself, or as individual skill folders with the skills CLI, which you can edit.
Skills vs MCP
An MCP server is a running program that exposes tools: query this database, create this ticket, search this API. A skill is static text and files that explain how to do a job. They solve different problems and work well together: an MCP server gives Claude the ability to call a service, and a skill tells Claude how and when to use it well.
If you are choosing between them, ask what is missing. If Claude cannot reach a system, you need an MCP server (or a CLI the skill can call). If Claude can reach it but does the job badly or inconsistently, you need a skill.
Which one should you write?
- A fact Claude must always know? Put it in
CLAUDE.md. - A procedure you keep re-explaining? Write a skill. Check it with our SKILL.md validator.
- Work that should run separately and report back? Use a subagent, possibly with skills preloaded.
- Access to an outside system? Connect an MCP server.
- A whole setup to share with a team? Package it as a plugin.