Claude Code is where skills started, and it still gets the most out of them: skills there can run scripts with full network access, take arguments, run in their own subagent and replace the custom slash commands you may already have. This guide covers the Claude Code side specifically. For the format in general and where skills come from, see Claude skills and agent skills.
We read Claude Code's documentation and command reference on 14 November 2026. The list above this guide is every skill in HostedMCP's catalog with the command to install it in Claude Code.
Where Claude Code looks
| Scope | Folder | Who gets it |
|---|---|---|
| Personal | ~/.claude/skills/<name>/SKILL.md |
You, in every project |
| Project | .claude/skills/<name>/SKILL.md |
Everyone who clones the repository |
| Nested | <subfolder>/.claude/skills/<name>/SKILL.md |
Work in that part of a monorepo |
| Plugin | <plugin>/skills/<name>/SKILL.md |
Anyone with the plugin, as /plugin:name |
| Organization | The managed settings folder | Everyone under that policy |
If two skills share a name, the organization's wins over yours, and yours over the project's. Skills you enable on claude.ai are synced into ~/.claude/skills/synced/.
Create one
mkdir -p .claude/skills/release-notes
Then write .claude/skills/release-notes/SKILL.md:
---
name: release-notes
description: Write release notes from merged pull requests. Use when the user asks for release notes, a changelog entry or a summary of what shipped.
---
1. List pull requests merged since the last tag with `gh pr list --state merged`.
2. Group them under Added, Changed and Fixed.
3. Write one plain sentence per change, for users rather than developers.
That is a working skill. Claude Code picks it up without a restart. Ask for release notes and it will use it, or type /release-notes.
Keep SKILL.md under 500 lines and move long material into files next to it that the instructions point to. Anthropic's skill-creator plugin, /plugin install skill-creator@claude-plugins-official, can test a skill with and without it loaded and tune its description.
Invoke it
| Frontmatter | You can type /name |
Claude uses it on its own | Description in context |
|---|---|---|---|
| Default | Yes | Yes | Always |
disable-model-invocation: true |
Yes | No | No |
user-invocable: false |
No | Yes | Always |
/name only works at the start of a message, and up to six can be chained. Skills accept arguments: $ARGUMENTS in the body becomes whatever you type after the name, and argument-hint shows a hint in the menu.
Some other fields Claude Code adds to the standard: when_to_use for a separate trigger description, model and effort to run a skill on a different model or effort level, allowed-tools and disallowed-tools, and context: fork to run the skill in a subagent so its work does not fill your main conversation.
Custom slash commands have merged into skills. .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both create /deploy, and old command files keep working.
The skills that ship with Claude Code
Claude Code's command reference marks 20 commands as skills. These are the ones you will meet most:
| Skill | What it does |
|---|---|
/code-review |
Reviews a diff, branch or pull request for bugs at an effort level you choose, and can fix or comment |
/simplify |
Four review agents look for reuse, simplification, efficiency and abstraction problems, then apply fixes |
/batch |
Splits a large change into 5 to 30 units and runs one background agent per unit |
/debug |
Turns on debug logging for the session and reads the log |
/loop |
Runs a prompt again on an interval, or paces itself |
/run and /verify |
Launch the app and check that a change actually works |
/doctor |
Checks the installation, settings, extensions and CLAUDE.md |
/fewer-permission-prompts |
Turns read-only commands you approve often into an allowlist |
/update-config |
Edits settings.json from a plain description |
/claude-api |
Loads current Claude API and SDK reference |
/dataviz |
Chart and dashboard guidance with a palette validator |
/init, which writes a CLAUDE.md for a new project, is a built-in command rather than a skill, and /review is an alias of /code-review. Claude Code does not ship Anthropic's document skills for Word, PDF, PowerPoint and Excel; install those from the anthropics/skills repository.
What skills cost
Every skill Claude can call on its own puts its description in context on every turn. Claude Code caps this in two ways:
- Each description, with
when_to_use, is cut at 1,536 characters in the listing. - All descriptions together share a budget of about 1% of the model's context window. Past it, Claude Code drops descriptions starting with the least used skills and keeps only their names.
Once used, a skill's content stays in the conversation. After an automatic compaction, the most recent use of each skill is re-attached, up to 5,000 tokens per skill and 25,000 in total.
/skills lists your skills and can sort them by token cost; /skill-doctor reports each skill's context cost and how often it is used. A skill you only ever call by hand should have disable-model-invocation: true, which removes its description from context entirely.
Skill, hook, subagent, MCP or plugin
| You want | Use | Why |
|---|---|---|
| Facts Claude needs every session | CLAUDE.md | Loads every time; keep it under 200 lines |
| Know-how for some tasks | A skill | Loads only when relevant |
| Something to happen every time | A hook | Deterministic, no context cost; "put guardrails in hooks" |
| Heavy work kept out of your context | A subagent | Separate context, returns a summary |
| Tools or data from another system | An MCP server | A skill can teach Claude how to use it |
| All of the above for a team | A plugin | Packages them as one install |
The pairing Anthropic documents most is a skill plus an MCP server: the server provides the tools, the skill provides the judgment about when and how to use them.
Safety
Skills in Claude Code can run any script with your user's rights and full network access. Read the scripts of any skill you did not write. allowed-tools lets a skill use tools without asking you; it narrows nothing. In a project, review .claude/skills/ in pull requests the way you would review code. Organizations can lock this down with managed settings, including allowing only skills from approved plugin marketplaces.
When it does not work
| What you see | Cause | Fix |
|---|---|---|
| Claude ignores the skill | Description does not match your wording | Rewrite it; add when_to_use |
/name not found |
Folder name and name differ, or a new top-level folder |
Match them; run /reload-skills |
| A plugin skill has a different name | Plugin skills are namespaced | Type /plugin:name |
| Some skills lost their descriptions | Listing budget exceeded | Remove unused skills or set disable-model-invocation: true |
| Your skill is shadowed | An organization or personal skill with the same name | Rename yours |