Claude Code slash commands: built-ins and custom ones

Built-in commands from the official table, the skills merge, and three files checked with claude plugin validate.

By David · Published October 31, 2026 · 12 min read

What Claude Code slash commands are

A slash command is recognized only at the start of your message. Text after the name is its arguments. The commands page, checked October 31, 2026, says you can chain up to six skills in one message at the terminal prompt. Type / to see what this session can run.

Three kinds show up in that table.

• Built-in command. The behavior is coded into the CLI. /help and /compact are this kind.
• Bundled skill. A prompt handed to Claude, the same shape as a skill you write. The page marks these rows Skill. /debug and /code-review are this kind.
• Your own file. A SKILL.md you save, or an older file in .claude/commands/. You type /name for either one.

The skills page says custom commands have been merged into skills. The old URL https://code.claude.com/docs/en/slash-commands now serves that same page. I downloaded both markdown URLs on October 31, 2026. sha256sum printed one hash, and wc -c printed 118,879 bytes for each. The title on the file is “Extend Claude with skills”.

d2a5d07c0d1cef1d1edcf30f38f6b2712ff70ac050e3d50601cd3dadb42f9097  skills.md
d2a5d07c0d1cef1d1edcf30f38f6b2712ff70ac050e3d50601cd3dadb42f9097  slash-commands.md

Your old .claude/commands/ files keep working. Skills add a folder for supporting files, frontmatter for who invokes them, and the option for Claude to load them when the description matches. The same page says to prefer a skill for new work.

Built-in commands stay a separate list. A skill you name usage replaces /usage in a local terminal, and the alias /cost still runs the built-in. I did not try that replacement.

Claude Code commands cheat sheet

On October 31, 2026 I counted every data row in the All commands table. The file was https://code.claude.com/docs/en/commands.md, 57,789 bytes, sha256 dbeb2b14ca26b6684ea381cd841253b11b299b48eb1eb6a4869a0f8d8ec86831.

The table has 121 rows. 20 rows open with the Skill mark. 1 row, /deep-research, opens with the Workflow mark. The other 100 rows open with neither mark. Three of those 100 say Removed: /pr-comments, /ultraplan, and /vim. I left those three out of the list below. /review is its own row, and the text calls it an alias of /code-review, which is marked Skill. /plugin-authoring does not open with the Skill mark. The same cell says it is a skill from a built-in plugin.

The 21 rows below are built-in commands from that table. The one-line purpose is shortened from the page. This is not a third-party list. The downloadable copy is the Claude Code cheat sheet.

The skills page names six bundled skills in one sentence: /doctor, /code-review, /batch, /debug, /loop, and /claude-api. Each of those is marked Skill in the commands table. /checkup is an alias of /doctor. /review is an alias of /code-review. They are prompts. They are not the fixed logic behind /help.

/mcp manages servers you have already added. Add an MCP server to Claude Code is the command that writes the config, and the three scopes it lands in.

The page says the menu does not show every command to every user. /desktop is one example: macOS or x64 Windows, and a Claude subscription. I did not open a session, so the 121 rows are the document, not a menu I saw.

How to make Claude Code custom commands

Save a new command as a skill. The folder name is the command you type, unless you set name.

When two definitions share a name, the skills page ranks them. Enterprise wins over personal. Personal wins over project. A skill wins over a file in .claude/commands/. A plugin skill does not replace those, because it keeps the plugin prefix. A nested skill under a subfolder stays out of the launch set until Claude reads a file there. You can also type /apps/web:deploy for a skill at apps/web/.claude/skills/deploy/SKILL.md.

/reload-skills re-scans skill and command directories during a session. I did not run it inside a session.

Arguments, shell output, and allowed tools

$ARGUMENTS is everything after the command name. $0 is the first argument. $1 is the second. $ARGUMENTS[0] is the same as $0. Wrap a multi-word value in quotes if it should stay one argument. The skills page’s example is /my-skill "hello world" second, which sets $0 to hello world and $1 to second.

An indexed placeholder with no argument stays in the text as you wrote it. It does not become an empty string. That is why /write-test tells Claude to stop if $0 or $1 is still in the instructions it receives. A named placeholder from the arguments field is different: a missing one expands to empty. These three files do not set arguments.

argument-hint is the autocomplete hint, such as [file] [behavior]. It is not the parser. The page says Claude Code ignores a frontmatter field it does not recognize, and does not report an error. The validator run below matched that for a field named colour.

A line that starts with ! and a command in backticks runs in the shell first. Claude receives the output, not the command. The page calls this dynamic context injection. The command is not scanned a second time, so the output cannot smuggle another shell line. The ! has to be at the start of the line or right after whitespace. Set disableSkillShellExecution to true and each of those lines becomes a fixed notice instead of running. I did not run the git lines in the starter files. The validator does not run them either.

allowed-tools pre-approves the listed tools for the turn that invokes the skill. The grant ends when you send the next message. It does not remove other tools. The page says a project skill’s grant applies even in a folder you have not trusted, so read allowed-tools on skills you clone before you run Claude Code there. /standup names Bash(git log *) and Bash(git status *). /ship-check adds Bash(git diff *), Read, and Grep. /write-test names Read, Grep, and Glob, and no shell. I did not watch a permission prompt.

Claude Code slash commands examples

Three files for a web project. One is the older command layout. Two are skills. Download the Claude Code cheat sheet, then the three files: standup.md, ship-check/SKILL.md, and write-test/SKILL.md. The same files are on Templates.

/standup summarizes commits since yesterday. The shell line is git log --since=yesterday. /ship-check is manual only, takes one argument string, and names the tools it wants pre-approved. /write-test takes a file and a behavior as $0 and $1.

---
description: Summarize commits since yesterday into a short standup note. Use when you want status from git, not a code change.
argument-hint: "[optional focus]"
disable-model-invocation: true
allowed-tools: Bash(git log *) Bash(git status *)
---

Write a standup note from the git output below. Do not change files. Do not commit.

If the user passed a focus, keep the note on that: $ARGUMENTS

Group the commits into shipped, still open, and blocked. If the log is empty, say there were no commits since yesterday.

Recent commits:
!`git log --since=yesterday --pretty=format:%h %ad %s --date=short`

Working tree:
!`git status --short`

---
name: ship-check
description: Check a web project before a deploy. Use when the user runs /ship-check. Does not deploy and does not commit.
argument-hint: "[what changed]"
disable-model-invocation: true
allowed-tools: Bash(git status *) Bash(git diff *) Read Grep
---

Check whether this project is ready to ship. Do not deploy. Do not commit. Do not push.

What the user said changed: $ARGUMENTS

1. Read the status and the diff stat below.
2. Name problems you can see in that diff: a behavior change with no test, a secret, or a broken path in a file you can read.
3. If you cannot see a problem, say what you checked and that you found nothing.

Status:
!`git status --short`

Diff stat:
!`git diff --stat`

---
name: write-test
description: Write one test for a named file and a named behavior. Use when the user runs /write-test with a path and the behavior to cover.
argument-hint: "[file] [behavior]"
disable-model-invocation: true
allowed-tools: Read Grep Glob
---

Write one test. Do not change production code. Do not invent a test runner.

File: $0
Behavior to cover: $1

1. Read $0.
2. Find the test file that already covers that module. If none exists, say so and name the path you would add. Do not create a new runner.
3. Add one test for the behavior in $1. Match the runner and the file layout the repo already uses.
4. If the instructions you receive still contain the characters $0 or $1, that argument was not passed. Stop and say which one is missing.

Copy the .claude folder into the project. Then run the validator. Do not treat a passing validator as a session. These files were not executed by the model.

npm i -g @anthropic-ai/claude-code --prefix /home/ubuntu/.nvm/versions/node/v22.23.4

2.1.316 (Claude Code)

Validating components in: /workspace/public/templates/claude-code-commands/.claude/skills

✔ Validation passed

{
  "success": true,
  "strict": false,
  "target": "/workspace/public/templates/claude-code-commands/.claude/skills",
  "manifest": null,
  "contents": [],
  "advice": [],
  "directory": null
}

Validating components in: /workspace/public/templates/claude-code-commands/.claude/commands

✔ Validation passed

Validating components in: /tmp/cc-pub/unknown/skills

✔ Validation passed

Validating components in: /tmp/cc-pub/broken/skills

Validating skill: /tmp/cc-pub/broken/skills/ship-check/SKILL.md

✘ Found 1 error:

  ❯ frontmatter: YAML frontmatter failed to parse: YAML Parse error: Unexpected character. At runtime this skill loads with empty metadata (all frontmatter fields silently dropped).

✘ Validation failed

{
  "success": false,
  "strict": false,
  "target": "/tmp/cc-pub/broken/skills",
  "manifest": null,
  "contents": [
    {
      "file": "/tmp/cc-pub/broken/skills/ship-check/SKILL.md",
      "type": "skill",
      "errors": [
        {
          "path": "frontmatter",
          "message": "YAML frontmatter failed to parse: YAML Parse error: Unexpected character. At runtime this skill loads with empty metadata (all frontmatter fields silently dropped).",
          "code": null
        }
      ],
      "warnings": [],
      "notes": [],
      "gatingHooks": []
    }
  ],
  "advice": [],
  "directory": null
}

Slash command, skill, subagent, or CLAUDE.md

Use the file that matches the job. The skills page, the subagent page, and the memory page are the sources. Checked dates are on those posts and in Sources below.

How to write a Claude MD file is the standing file: where it loads, and a template. Claude Code subagents is the worker with its own window. A slash command does not replace either one. A skill can set context: fork when that one command should run in a subagent. The three files here do not set context.

Claude Code slash commands best practices

A few limits from the skills page, not a scored list.

• Put a new command in .claude/skills/name/SKILL.md. Leave old .claude/commands/ files where they are. They still run.
• Set disable-model-invocation: true on anything with a side effect. Deploy and commit are the page’s examples. /ship-check in this pack follows that.
• Write description as the case you want matched. Put the key use case first. The listing caps that text at 1,536 characters.
• Keep SKILL.md under 500 lines. The page says to move long reference into a neighboring file.
• Read allowed-tools before you run a skill you did not write. The grant applies even in a folder you have not trusted, on the page’s own warning.
• Run claude plugin validate and claude plugin validate --strict on .claude/skills. On 2.1.316 an unknown field was silent, and an unclosed bracket was silent. An unclosed quote was not.

More on this topic lives under AI agents and tools.