Skip to content

This page was translated from the Korean original with AI and may contain errors.

Chapter 9. Skills and Extensions ​

Required Read this to follow this stage.

One-line summary: Turn instructions you keep repeating into a skill. Side tasks go to subagents, outside services go through MCP, and bundled installs are plugins. At first, one skill is enough.

The screens in this chapter come from running Claude Code 2.1.318 on Linux on November 6, 2026.

Making a Skill ​

A skill (Skill) is a feature that adds a work procedure for Claude with a single SKILL.md file. If you keep pasting the same instructions into the chat, it is worth making one.[1] Think of it as a work manual you call by name.

Project skills go in .claude/skills/<skill-name>/SKILL.md.[1:1] Leave with /exit and move to the practice folder in the terminal. On Mac and Linux, paste the whole block below (view as a file: ch9-skill-meeting-notes.txt).

bash
mkdir -p .claude/skills/meeting-notes
cat > .claude/skills/meeting-notes/SKILL.md <<'EOF'
---
description: Turns meeting notes into meeting minutes that follow the rules. Use it for requests to organize meeting notes.
---

Read the $ARGUMENTS file and make meeting minutes following the CLAUDE.md rules.
Name the file `minutes-date.md` and save it in the notes folder.
When done, report the saved file name and the number of to-dos in one line.
EOF

If it has not run yet after pasting, press Enter once more. If nothing appears, it worked. Check with ls .claude/skills/meeting-notes that SKILL.md was created.

In Windows PowerShell, run mkdir .claude\skills\meeting-notes, then open Notepad with notepad .claude\skills\meeting-notes\SKILL.md. Leave out the line starting with cat and the EOF line, and put in only what is between them. If it asks whether to create a new file, click Yes, save with Ctrl + S, and close it. We did not test this method.

The file has two parts: the settings between the --- lines and the instructions below them.[1:2]

  • The folder name meeting-notes becomes the command name you type.[1:3]
  • description is what Claude uses to decide when to load this skill.[1:4]
  • $ARGUMENTS is a blank. The text you pass when calling the skill goes here.[2] In the hands-on, the file path written after the command went in, and the skill organized that note. So you can use the same skill to organize other notes too.
text
Weak description:   meeting
Better description: Turns meeting notes into meeting minutes that follow the rules. Use it for requests to organize meeting notes.

The better example says both what it does and when to use it. This makes it easier for Claude to notice the right moment.

Now start claude in the terminal, type /, and look for it in the list. To call it yourself, type this:

text
/meeting-notes inbox/meeting-2026-10-27-weekly.txt

When we called the same skill with claude -p from Chapter 10, notes/minutes-2026-10-27.md was created and Claude answered like this:

text
Saved to `notes/minutes-2026-10-27.md`, with 5 to-dos.

Following the CLAUDE.md rules from Chapter 8, it produced three sections: decisions, to-dos, and next meeting (see the result). The skill body loads only when it is used, so it costs almost nothing the rest of the time.[1:5]

Quick check

You made a skill at `.claude/skills/meeting-notes/SKILL.md`. Which command calls it directly?

When a Skill Does Not Work ​

Sometimes a mistake does not cause an error. If the skill is not in the list or behaves oddly, check in this order:

  1. Is the folder name the same as the command you typed?
  2. Is it located at .claude/skills/meeting-notes/SKILL.md?
  3. Is the very first line of the file ---? The settings are read only when it is the first line.[3]
  4. Are the setting field names spelled correctly? If not, they are ignored without an error.[3:1]
  5. Did you start claude in the practice folder? Check with pwd.

Changes you make during a session take effect without a restart. But if you create a new skill folder that did not exist when the session started, run /reload-skills.[3:2]

Quick check

It seems the skill's settings are not being read. Where is a good place to look first?

Subagents ​

A subagent (Subagent) is a specialist helper that takes on a specific kind of work. It works in its own context window and returns only a summary. So search results and logs do not clutter your conversation.[4] It is like giving research to an assistant and hearing only the conclusion.

For example, imagine checking five meeting minutes one by one against the original notes. If you ask directly, the contents of ten files of notes and minutes pile up in your conversation. If you hand it to a subagent, the subagent reads the files, and only a few lines of conclusions come back to your conversation, like "Minutes #3 has a different deadline from the note."

The easiest way to make one is to ask Claude.[4:1] Just say something like "Make a subagent in this project that checks whether the meeting minutes match the original notes." If it cannot find the new one, try restarting Claude Code.[4:2] This guide did not actually make a subagent.

Older articles may tell you to make one in the /agents screen. That wizard is gone now. Typing /agents only shows a message telling you to ask Claude or edit the files yourself.[4:3]

A subagent's requests count toward the same usage limits as your conversation.[4:4] For small tasks, it is better to just ask directly.

Quick check

In recent versions of Claude Code, what is the easiest way to make a subagent?

MCP ​

MCP (Model Context Protocol) lets Claude Code do things beyond its basic tools. Examples include searching a list of to-dos and bug reports (an issue tracker), looking things up in a store of data (a database), and controlling a web browser.[5] Think of it as plugging a cable into Claude that leads to an outside service.

There are two places to type commands. You type claude mcp add in the terminal (shell) before starting claude. You type /mcp inside a session to check servers you have already added.[5:1] We tried adding, checking, and removing the example server from the official docs.

bash
claude mcp add --transport http claude-code-docs https://code.claude.com/docs/mcp
claude mcp list
claude mcp remove claude-code-docs

Let's read the first line piece by piece.

  • claude mcp add: Adds an MCP server.
  • --transport http: Means it is a remote server connected over HTTP.
  • claude-code-docs: The server name you give it.
  • https://…: The server address.

list shows the list, and after remove you write the name of the server to delete. Running the three commands prints these lines in order:

text
Added HTTP MCP server claude-code-docs with URL: https://code.claude.com/docs/mcp to local config
claude-code-docs: https://code.claude.com/docs/mcp (HTTP) - ✔ Connected
Removed MCP server "claude-code-docs" from local config

✔ Connected means it is connected. The default scope local means only you use it, and only in this project.[5:2]

Connect only servers you can trust. Servers that fetch outside content can expose you to prompt injection (a trick that tries to control an AI with instructions hidden in text).[6] Do not add a server when you do not know who made it.[7]

Quick check

Where should you type the `claude mcp add` command?

Plugins ​

A plugin (Plugin) is a folder that installs skills, agents, hooks, MCP servers, and more as one bundle. A hook is a command that runs automatically at set moments; you will learn about it in Chapter 11. Everything above can be used on its own without a plugin. You only need a plugin when you want to use several of them as a bundle.[8]

For example, imagine bundling this chapter's meeting-minutes skill and the checking subagent into one plugin. A coworker would not need to make the two one by one. They would just install the one plugin.

Inside a session, type /plugin and you can choose plugins in the Discover tab.[8:1] The install command looks like /plugin install commit-commands@claude-plugins-official. If you type it in a session, it does not install right away; a details screen opens first.[9] This guide did not go as far as installing one.

Plugins run with your permissions. Whatever the source, an installed plugin can run code with your user permissions.[8:2] That is why you should read the details screen before installing.[9:1] Treat it like a program you run yourself.

An enabled plugin becomes part of every session, even sessions where you do not use it.[8:3] If you do not use it, disable or remove it in the Installed tab of /plugin.[9:2]

Quick check

What should you remember before installing a plugin?

What to Use When ​

WhenUse
You have rules to always followCLAUDE.md
You repeat the same procedureSkill
You hand off a big side taskSubagent
You use an outside serviceMCP
You want several things in one installPlugin
It must run every single timeHook (Chapter 11)

You do not need to install everything. First, ask yourself: Is anything bothering me right now? If not, you do not have to make anything yet.

Quick check

Every time you organize meeting notes, you paste the same steps into the chat. What should you make?

Try It Yourself ​

  1. Make the skill in the terminal. Start claude, type /, and check that meeting-notes appears.
  2. Run /meeting-notes inbox/meeting-2026-10-27-weekly.txt. Was a file created in notes? Does its content match the note?
  3. Leave with /exit and type the three claude mcp commands in the terminal in order. Check with claude mcp list that the server was removed.

Ask yourself: Could I organize other notes with the same skill? Is there an extension I really need right now?

Summary ​

  • A skill is made with a single .claude/skills/<name>/SKILL.md. The folder name is the command name.
  • If it does not work, check the folder name, the --- on the first line, and the spelling of the field names.
  • The /agents wizard is gone. Make subagents by asking Claude or by writing a file.
  • Type claude mcp add and similar commands in the shell, and /mcp inside a session.
  • Use only MCP servers and plugins you can trust. Plugins run with your permissions.

Glossary ​

TermOne-line meaning
SkillA work manual you call by name
SubagentA helper that handles side tasks in its own context
MCPA way to connect outside tools
PluginAn install unit that bundles several extensions

Finished this chapter?

Progress is saved only in this browser. See your progress on the full table of contents.

Next chapter: Chapter 10. One-Line Runs and Automation

References (9)
  1. Claude Code official docs, "Extend Claude with skills". "Create a SKILL.md file with instructions, and Claude adds it to its toolkit." / "Create a skill when you keep pasting the same instructions, checklist, or multi-step procedure into chat, or when a section of CLAUDE.md has grown into a procedure rather than a fact." / "Commit it so your team gets it too" / "Every skill needs a SKILL.md file with two parts: YAML frontmatter between --- markers that tells Claude when to use the skill, and markdown content with the instructions Claude follows when the skill runs." / "The directory name, or the frontmatter name when you set one, becomes the command you type, and the description helps Claude decide when to load the skill automatically." / "Unlike CLAUDE.md content, a skill's body loads only when it's used, so long reference material costs almost nothing until you need it." code.claude.com (checked on 2026-11-06) ↩︎

  2. Claude Code official docs, "Extend Claude with skills". "All arguments passed when invoking the skill." code.claude.com (checked on 2026-11-06) ↩︎

  3. Claude Code official docs, "Extend Claude with skills". "Claude Code reads the frontmatter only when the opening --- is the file's first line." / "Claude Code ignores a field it doesn't recognize without reporting an error." / "When you add, edit, or remove a skill under ~/.claude/skills/, the project .claude/skills/, or a .claude/skills/ inside an --add-dir directory, Claude Code picks up the change within the current session, without a restart." / "Claude Code isn't watching that directory yet, so run /reload-skills again after each later change there." code.claude.com (checked on 2026-11-06) ↩︎

  4. Claude Code official docs, "Create custom subagents". "Subagents are specialized AI assistants that handle specific types of tasks." / "Use one when a side task would flood your main conversation with search results, logs, or file contents you won't reference again: the subagent does that work in its own context and returns only the summary." / "Running /agents prints a reminder to ask Claude or edit .claude/agents/ and ~/.claude/agents/ directly." / "To create one, ask Claude to write it for you, or write the file yourself." / "If Claude can't find the new subagent, restart Claude Code and try again." / "It also sends its own requests, which count toward the same usage limits as your main conversation." code.claude.com (checked on 2026-11-06) ↩︎

  5. Claude Code official docs, "Connect to MCP servers". "lets Claude Code use tools beyond its built-in set, such as searching an issue tracker, querying a database, or controlling a web browser." / "Run this in your terminal, not inside a claude session: you're configuring the server before starting a conversation." / "Inside a claude session, use the /mcp command to check and manage servers you've already added." / "By default it registers the server at local scope: private to you, active only in the current project." code.claude.com (checked on 2026-11-06) ↩︎

  6. Claude Code official docs, "Connect Claude Code to tools via MCP". "Verify you trust each server before connecting it." / "Servers that fetch external content can expose you to prompt injection risk." code.claude.com (checked on 2026-11-06) ↩︎

  7. Anthropic does not security-check MCP servers for you either. Claude Code official docs, "Security". "Anthropic reviews connectors against its listing criteria before adding them to the Anthropic Directory, but does not security-audit or manage any MCP server." code.claude.com (checked on 2026-11-06) ↩︎

  8. Claude Code official docs, "Plugins overview". "A Claude Code plugin is a directory of skills, agents, hooks, MCP servers, or other components that Claude Code installs and loads as one unit." / "Skills, subagents, hooks, and MCP servers all work on their own, without a plugin." / "To try a plugin now, run /plugin in a Claude Code terminal session and install one from the Discover tab, which lists the plugins from your marketplaces." / "Whatever the tier, a plugin you install can run code with your user privileges." / "An enabled plugin is part of every session, not only the sessions where you use it." code.claude.com (checked on 2026-11-06) ↩︎

  9. Claude Code official docs, "Install and manage plugins". "/plugin install commit-commands@claude-plugins-official" / "In a session, this command doesn't install right away: it opens the /plugin panel on that plugin's details so you can review it and choose a scope first." / "A plugin can run hooks and MCP servers, so read the pane before you install." / "The Installed tab in /plugin lists your plugins with actions to enable, disable, update, or uninstall each one." code.claude.com (checked on 2026-11-06) ↩︎