Claude Code Slash Commands: The Built-in List and How to Write Your Own

Slash commands are shortcuts you type at the Claude Code prompt. Built-ins such as /compact, /model and /permissions control the session, and custom commands are skills you trigger by name. To write your own today, create .claude/skills/<name>/SKILL.md with disable-model-invocation: true and type /<name>; legacy .claude/commands/*.md files still work.

What slash commands are

There are three kinds:

• Built-in commands ship with Claude Code and control the session, settings and account.
• Custom commands are skills you invoke by name. They can come from your project, from your home directory, or from a legacy .claude/commands/ file.
• Plugin commands arrive with an installed plugin and are namespaced as /plugin-name:skill-name.

Type / at the prompt to open the menu, keep typing to filter it, and press Tab to complete a name. /help lists every available command, including your custom skills. The set changes between releases, so the menu on your machine beats any published list, this one included.

Built-in commands worth knowing

These are highlights from Anthropic's commands reference, grouped by job. The list isn't complete, and commands change between releases; type / or run /help to see every command on your version. Square brackets mark optional arguments and angle brackets mark required ones.

A few of these pay off on day one:

• /clear between unrelated tasks.
• /compact with instructions about what to keep, when context runs long.
• /rewind when an edit goes wrong, because it rolls back code and conversation together.

Managing memory and context covers /compact, /context and /memory in depth. Settings and permissions explains what /permissions edits.

Custom commands are skills now

Custom commands have been merged into skills. .claude/commands/*.md files still work as a legacy path, but .claude/skills/<name>/SKILL.md is canonical, and every skill can be invoked as /<name>. What makes a skill behave like a classic slash command is one frontmatter line, disable-model-invocation: true, which stops Claude from loading it on its own.

Project locations are committed and shared with your team; personal ones follow you into every repo. The command name comes from a different place in each format:

• A legacy command's name comes from its filename.
• A skill's command comes from its directory name. The name field only sets the label shown in skill listings, except in a plugin skill, where it replaces the command's last segment.

Monorepos can also keep .claude/skills/ directories further down the tree.

For new work, write skills. A skill is a folder, so it can bundle scripts and templates next to SKILL.md. If you later want a command to auto-load, you delete one line instead of moving files.

Anatomy of a command file

A command is YAML frontmatter plus a Markdown body. The frontmatter controls how it's invoked, and the body is the prompt Claude receives when you run it.

---
name: pr-summary
description: Summarize the diff against a base branch as a PR description.
argument-hint: [base-branch]
disable-model-invocation: true
allowed-tools: Bash(git *) Read
---
 
Summarize the changes on this branch against $0.
 
1. Run git diff $0...HEAD and read the changed files.
2. Write an imperative title and a body with Summary and Test plan sections.
3. Don't push or open anything; print the draft.

Save that as .claude/skills/pr-summary/SKILL.md and run /pr-summary main. The body runs in your main conversation, as if you had pasted it, unless you set context: fork.

The frontmatter fields that shape invocation:

The spellings are exact. The names use hyphens, not underscores (disable-model-invocation, user-invocable), and context: fork takes the literal value fork.

The legacy format is the same idea in a single file. A .claude/commands/<name>.md file takes its name from the filename, has optional frontmatter, and a body that reads $ARGUMENTS:

---
description: Review a pull request and list findings by severity.
argument-hint: [PR number]
allowed-tools: Bash(gh *) Read
---
 
Review pull request #$ARGUMENTS. Run gh pr diff $ARGUMENTS, then list blockers, should-fix items and nits with file:line for each.

To convert it, move the file to .claude/skills/<name>/SKILL.md and add disable-model-invocation: true.

Passing arguments

Whatever you type after the command name becomes arguments:

• $ARGUMENTS expands to the whole argument string as typed.
• $0, $1 and so on pick out single arguments, counting from zero.
• Values follow shell-style quoting, so /my-skill "hello world" second puts hello world in $0 and second in $1.
• An index with no matching argument is left in the text untouched.

Named arguments read better once there are two or more. Declare them in frontmatter and use them by name:

---
argument-hint: [issue] [branch]
arguments: [issue, branch]
disable-model-invocation: true
---
 
Fix issue $issue on branch $branch.

Names map to positions in order, so $issue takes the first argument and $branch the second. Named and positional placeholders differ in a few ways:

• A named argument with nothing passed expands to an empty string, while a positional one is left untouched.
• To write a literal dollar sign, escape it with a backslash, as in \$1.00.
• A single-argument hint such as argument-hint: [base-branch] still lands in $0.

Always tell the body what to do when no argument arrives. The commands in our library ask one focused question instead of guessing.

Run shell commands before Claude sees the prompt

A skill can inject live data before Claude reads it. Put !`git diff HEAD` in the body and the command runs when the skill loads. Its output replaces the placeholder, so Claude starts from the current diff instead of spending a turn fetching it. For multi-line commands, use a fenced code block marked with !.

Three substitutions help a skill find its surroundings:

• ${CLAUDE_SKILL_DIR} is the skill's own directory, which is useful for bundled scripts.
• ${CLAUDE_PROJECT_DIR} is the project root.
• ${CLAUDE_SESSION_ID} is the current session.

Injected commands run on your machine as you, so read them before adopting someone else's command. The skills security checklist lists what to look for.

Project vs personal commands, and namespacing

Put a command in the project when the team should share it and it depends on the repo, such as build steps, review checklists or release flows. Put it in ~/.claude/skills/ when it's your own habit and works anywhere.

Plugins solve the naming-collision problem, because plugin skills are always namespaced as /plugin-name:skill-name:

• A plugin named my-plugin with a skill hello runs as /my-plugin:hello, so two plugins can both ship a review command.
• The prefix comes from the name field in plugin.json.
• A plugin can still carry a legacy commands/ directory, but skills/ is canonical.

After editing a plugin, /reload-plugins picks up the change without a restart. Claude Code plugins covers building and installing them. MCP servers are managed separately through /mcp; see adding MCP servers to Claude Code.

Letting Claude invoke commands, and turning that off

By default a skill works both ways: you can type it, and Claude can load it through the Skill tool whenever its description matches the task. Two frontmatter flags split those paths:

disable-model-invocation: true is the right default for anything with side effects, such as committing, pushing, deploying or opening PRs, because you decide when it runs. It has one cost: a subagent can't preload such a skill through its skills field, because preloading counts as model invocation.

To change a skill's behavior without editing its file, add skillOverrides to .claude/settings.json:

{
  "skillOverrides": {
    "my-skill": "off",
    "another-skill": "user-invocable-only"
  }
}

Five ready-made commands from our library

If you'd rather start from working examples, these AgentsCamp commands cover loops most developers repeat daily. Each page shows the full file:

1. Review PR, run as /review-pr [PR number], reviews a pull request end to end. It:
   • pulls metadata, diff and CI status through gh;
   • checks out the branch and evaluates correctness, security and style against the PR's stated goal;
   • tags each finding blocker, should-fix or nit, with file:line and a fix;
   • ends with a verdict and never merges.
2. Commit inspects status, diffs and recent log. It stages only the files that belong to one logical change and writes a Conventional Commits message with an imperative subject of 72 characters or fewer. It flags secrets instead of staging them, and never pushes unless asked.
3. Create PR, run as /create-pr [base branch or notes], opens a pull request. It confirms the base branch, pushes if needed, and writes a title and a body with summary, changes, testing and risk. It opens the PR through gh and reports the URL, without merging.
4. Fix Failing Test, run as /fix-failing-test [test name or path], reproduces the failure and traces it to the code under test. Before editing, it states whether the test or the code is wrong. Then it applies the smallest fix to the correct side and reruns the suite, never weakening an assertion.
5. Explain Error takes an error message or stack trace. It finds the first in-repo frame, names the root cause one level below the message, and confirms it with read-only checks. It reports a specific fix, or ranked candidates, and changes no files.

Copy a file from its page, or install from the terminal. The command below installs into your project's .claude/ directory by default; add -g to install into ~/.claude/.

npx agentscamp add commands/review-pr

To build your own, Create Slash Command scaffolds a new command file from a one-line description.

When a command should become a skill or a subagent

Commands are the manual end of Claude Code's extension model. These signals say it's time to move:

The deciding questions are who pulls the trigger and whether the work needs its own context window. Skills vs agents vs commands works through that decision with examples. When you are ready, Create Skill and Create Subagent scaffold the next step.

Next steps

Start with one command for the prompt you retype most. Commit it under .claude/skills/, and add an argument-hint so teammates can see how to call it. Then:

• Read Claude Code skills best practices for descriptions and bodies that hold up as the library grows.
• Use Claude skills not working? when a command doesn't appear in the menu.
• Browse 25 Claude Code tips for the shortcuts around the command menu.