Claude Code Commands: A Cheat Sheet That Says Where It Got Each Line

Steps

1. 01

   Print the authoritative list for your build first

   Claude Code ships new subcommands and retires old ones between releases, and plugins add more, so the only list that is definitely right is the one your own install prints. Run these three before trusting any cheat sheet, including this one. claude --help lists the CLI surface; /help inside a session lists the slash commands that are actually registered for you; claude doctor tells you which build you are on and where it lives.

   # what version am I actually running?
   claude --version
   
   # the full CLI surface for that build
   claude --help
   
   # where it is installed, and whether anything is broken
   claude doctor

   Copy
2. 02

   The difference nobody spells out: CLI subcommands vs slash commands

   There are two separate command surfaces and they are not interchangeable. CLI subcommands are typed in your shell before Claude Code starts — claude mcp list, claude doctor, claude update. Slash commands are typed inside a running session — /compact, /clear, /memory — and they do not work from the shell. A cheat sheet that mixes them into one list is the single most common reason a command "does not work": you are typing a session command at a shell prompt, or the reverse. The two tables below are deliberately kept apart.

   # CLI subcommand — from your shell, Claude Code not running
   claude mcp list
   
   # slash command — typed at the prompt INSIDE a session
   claude
   > /compact

   Copy
3. 03

   One-shot mode: the flag combination worth memorising

   The -p/--print flag runs a prompt without entering the interactive shell and prints the result to stdout, which is what makes Claude Code scriptable. It combines with --output-format to get JSON instead of prose, and with --max-budget-usd to cap what a single run can spend. All four flags below are verbatim from claude --help on v2.1.260. Note the warning the help itself carries about -p: the workspace trust dialog is skipped in non-interactive mode, so only use it in directories you trust.

   # print and exit
   claude -p "Explain what this repo does"
   
   # machine-readable result
   claude -p "List every TODO in src/" --output-format json
   
   # cap the spend of this one run
   claude -p "Refactor the parser" --max-budget-usd 2
   
   # stream events as they happen
   claude -p "Fix the failing test" --output-format stream-json --include-partial-messages

   Copy
4. 04

   When something is broken, start it with everything switched off

   Two flags exist purely to answer "is it Claude Code, or is it my configuration?". --safe-mode starts a session with CLAUDE.md, skills, plugins, hooks, MCP servers, custom commands and agents all disabled, while auth, models, built-in tools and permissions keep working; admin-managed policy settings still apply. --bare goes further and skips hooks, LSP, plugin sync, auto-memory, keychain reads and CLAUDE.md auto-discovery entirely. If a session works in safe mode and fails without it, the cause is one of your customisations, and you have just cut the search space to your own config directory.

   # everything you configured, off
   claude --safe-mode
   
   # minimal mode: no hooks, no LSP, no CLAUDE.md discovery
   claude --bare
   
   # only the MCP servers in this file, ignoring every other MCP config
   claude --strict-mcp-config --mcp-config ./mcp.json

   Copy
5. 05

   Where settings live, and how to override them for one run

   Claude Code reads settings from three scopes: user (~/.claude/settings.json), project (.claude/settings.json, meant to be committed and shared) and local (.claude/settings.local.json, personal and normally git-ignored). Two flags let you control that from the command line without editing any file. --setting-sources takes a comma-separated list of which scopes to load at all, and --settings takes either a path to a JSON file or a raw JSON string layered on top. This is the pair you want in CI, where project settings should apply but a developer’s local overrides should not.

   # load only the committed project settings — ignore user and local
   claude --setting-sources project
   
   # layer an extra settings file on top
   claude --settings ./ci-settings.json
   
   # or pass the JSON inline
   claude --settings '{"model":"sonnet"}'

   Copy
6. 06

   Running Claude Code on Windows, natively or through WSL

   Both work. v2.1.260 publishes native Windows builds — its package.json lists @anthropic-ai/claude-code-win32-x64 and @anthropic-ai/claude-code-win32-arm64 as optional dependencies next to the macOS and Linux ones — so a plain npm install -g in PowerShell gets you a native binary and no Linux layer is required. WSL is still supported and is the better choice when your project already lives on a Linux filesystem, because the agent runs shell commands constantly and WSL avoids path and line-ending friction. The one combination to avoid is running Claude Code inside WSL while pointing it at a Windows-side npm installation under /mnt/c: Claude Code detects that setup specifically and complains about it, and it is the most common broken Windows install.

   # PowerShell — native Windows build
   npm install -g @anthropic-ai/claude-code
   claude --version
   
   # inside WSL — install with the LINUX npm, not the one under /mnt/c
   which npm      # expect /usr/bin/npm or ~/.nvm/..., NOT /mnt/c/...
   npm install -g @anthropic-ai/claude-code

   Copy
7. 07

   Uninstalling, and what stays behind

   Removing the npm package does not remove your data, which is usually what people actually want gone. The package removal is one command. Your account metadata lives in ~/.claude.json and your projects, transcripts, skills and settings live under ~/.claude/ — claude doctor prints that projects directory, so you can confirm it before deleting anything. The OAuth token itself is not in either place: it is in your OS keychain under "Claude Safe Storage". If you only want one project’s history gone rather than the whole install, claude project purge deletes the transcripts, tasks, file history and config entry for a single project and leaves everything else alone.

   # remove the CLI
   npm uninstall -g @anthropic-ai/claude-code
   
   # sign out first if you want the stored credential gone
   claude auth logout
   
   # just one project’s state, not the whole install
   claude project purge
   
   # then, only if you really want a clean slate:
   # rm -rf ~/.claude ~/.claude.json

   Copy

Troubleshooting

command not found: claude, right after npm install -g succeeded

npm's global bin directory is not on your PATH, so this is a shell problem and not a Claude Code problem. Find the directory with npm config get prefix, add /bin to it, and put that on your PATH: export PATH="$(npm config get prefix)/bin:$PATH" in ~/.zshrc or ~/.bashrc, then source it. On macOS with Homebrew-managed Node the prefix is usually /opt/homebrew. Confirm with which claude before trying anything else.

npm ERR! EACCES permission denied on the global install

You are writing to a system-owned directory. Do not fix this with sudo — sudo npm install -g leaves root-owned files in your npm tree and breaks later installs. Point npm at a directory you own instead: mkdir -p ~/.npm-global && npm config set prefix ~/.npm-global, then add export PATH="$HOME/.npm-global/bin:$PATH" to your shell profile. Using nvm or Volta avoids the whole class of problem because they install Node under your home directory.

npm ERR! engine: requires node >=22

The floor moved. v2.1.183 required Node >=18; v2.1.260 declares engines.node ">=22.0.0", so a machine that installed Claude Code earlier in 2026 will now fail on upgrade rather than on first install. Check with node --version, then nvm install --lts && nvm use --lts, and re-run the install.

A slash command "does not exist" — /vim, /todos, /output-style, /agents

Check them against your own build before assuming it is broken. In v2.1.260, /agents is still registered but its own description begins with "(removed)". None of /vim, /todos or /output-style is in the v2.1.260 command registry: /todos does not occur in the binary at all, and the only occurrences of the literal strings "/vim" and "/output-style" are inside unrelated text (a ripgrep option description and a docs URL). Treat all three as gone rather than as something you are typing wrong. This is also why an undated cheat sheet is worse than none: we fetched the one currently ranking on page one for this query on 2026-09-04, and it still lists --think, /save and a CLAUDE_API_KEY environment variable, with no "last updated" anywhere in its 103,889 bytes. Run /help; that list is the only one that is current for you.

An MCP server shows as failed to start, or its tools never appear

Work through it in this order. claude mcp get <name> health-checks an approved server and tells you what it is doing. If the server came from a project .mcp.json, it needs explicit approval before its tools are exposed — if you rejected it by accident, run claude mcp reset-project-choices in the project directory and restart claude to get the prompt back. If it is a stdio server, run the command by hand in the same shell first: most "failed to start" cases are just npx not finding the package. If it is an HTTP or SSE server that needs OAuth, claude mcp login <name> is a separate step from adding it.

Claude Code behaves differently in one repo than everywhere else

Something in that repo’s configuration is loading. Start with claude --safe-mode, which disables CLAUDE.md, skills, plugins, hooks, MCP servers and custom commands in one go. If the odd behaviour disappears, narrow it by scope: claude --setting-sources user ignores the project and local settings files, and --strict-mcp-config ignores every MCP config except the one you pass. Working outward from safe mode is much faster than reading four settings files.

Claude Code is not working at all — how do I tell if it is me?

Check the service first: status.claude.com is the status page for the API that Claude Code calls, and downdetector-style reports will not tell you anything it does not. If the service is fine, run three commands. claude --version proves the binary runs. claude doctor prints the install path, the build, and whether settings and search are healthy. claude auth status prints JSON with a loggedIn field — an expired login shows up as a refusal the moment you send a prompt, not at startup, which is why it is so often mistaken for an outage.