Claude Code Cheat Sheet: Commands, Shortcuts, and Prompts

This is the 5-minute Claude Code tutorial. You need a terminal, a project folder, and a Claude subscription (Pro, Max, Ultra, Team, or Enterprise) or a Claude Console account with API credits.

The native installer is recommended because it updates itself. Homebrew and WinGet installs don't, so upgrade those now and then. npm also works (npm install -g @anthropic-ai/claude-code, Node.js 22+), but don't use sudo.

On first launch, Claude Code opens your browser to log in. Then ask plain-English questions like "what does this project do?" or "explain the folder structure".

Not a terminal person? Claude Code also runs as a desktop app, in VS Code and JetBrains, and at claude.ai/code. This cheat sheet covers the terminal version.

Slash commands are typed inside a running session. Type / to open the menu and keep typing to filter it. Some are built-in commands. Others, like /code-review, /loop, and /doctor, are bundled skills: prompt-based playbooks that ship with Claude Code. The full list of 100+ is in the commands reference.

These run from your shell, before a session starts.

• claude "fix the build error": start a session with a first prompt.
• claude -c (--continue): load the most recent conversation in this folder.
• claude -r "auth-refactor" (--resume): resume by ID or name. No argument opens a picker.
• claude -n "auth-refactor" (--name): name a new session.
• claude -w feature-auth (--worktree): start in an isolated git worktree so parallel sessions don't collide.
• --model sonnet: pick a model by alias (sonnet, opus, haiku, fable) or full name.
• --permission-mode plan: start in a given permission mode (see below).
• --add-dir ../lib: let Claude read and edit another folder too.
• --allowedTools "Bash(git diff *)" "Read": pre-approve tools. --disallowedTools blocks them.
• claude update, claude -v, claude auth status: update, check the version, check your login.

--dangerously-skip-permissions turns off every permission check. Anthropic says to use it only in isolated containers or VMs, not on your laptop.

Press ? on an empty prompt to see the shortcuts for your terminal. The ones that matter:

• Esc: stop Claude mid-turn. Work so far is kept, so you can redirect.
• Esc Esc on an empty prompt: open the rewind menu.
• Shift+Tab: cycle permission modes.
• Ctrl+C: interrupt or clear input. Ctrl+D twice exits.
• Ctrl+B: send a running command or agent to the background.
• @: file autocomplete, which puts the file into your message.
• ! at the start: shell mode. Runs the command and shows Claude the output.
• Ctrl+V (Cmd+V in iTerm2, Alt+V on Windows): paste an image, like a screenshot of an error.
• Ctrl+G: open your prompt, or a proposed plan, in your text editor.
• Ctrl+R: search prompt history. Ctrl+O: transcript with full tool calls.
• Option+P / Alt+P: switch model without clearing your prompt.
• New line without sending: \ then Enter, or Ctrl+J anywhere. Shift+Enter works natively in iTerm2, WezTerm, Ghostty, Kitty, Warp, Apple Terminal, and Windows Terminal. In VS Code, Cursor, Alacritty, or Zed, run /terminal-setup once.

On macOS, Alt shortcuts need "Option as Meta" turned on in your terminal.

A permission mode decides what Claude can do without asking. Shift+Tab cycles them, and the status bar shows the current one.

• default (shown as Manual): reads run freely, everything else asks.
• acceptEdits: file edits and common file commands (mkdir, mv, cp) run without asking.
• plan: Claude researches and writes a plan but can't edit your source until you approve.
• auto: actions run without prompts, but a separate classifier model reviews each one and blocks risky ones. Since v2.1.283 this is the built-in starting mode in the terminal and VS Code.
• dontAsk: only pre-approved tools run; anything else is denied. Built for CI.
• bypassPermissions: no checks. Containers and VMs only.

Plan mode is the best habit for anything bigger than a one-line fix. Enter it with Shift+Tab or /plan. When the plan is ready, approve it (with auto mode or manual edit review) or choose "No, keep planning" and give feedback. Ctrl+G lets you edit the plan first.

Note that claude -p runs start in Manual mode, so anything you didn't pre-approve gets denied instead of prompting.

CLAUDE.md is a Markdown file Claude loads into every session. It's guidance, not enforced configuration, so keep it short and specific. Anthropic recommends under 200 lines.

• ~/.claude/CLAUDE.md: your preferences for all projects.
• ./CLAUDE.md or ./.claude/CLAUDE.md: the project file. Commit it.
• ./CLAUDE.local.md: your personal notes for this project. Add it to .gitignore.
• .claude/rules/*.md: rules with a paths: filter that load only for matching files.
• @path/to/file inside CLAUDE.md imports another file.
• If the repo has an AGENTS.md (the file Codex and other agents use) and no CLAUDE.md, Claude Code reads AGENTS.md.
• Auto memory is separate: Claude keeps its own notes in ~/.claude/projects/<project>/memory/. Manage it with /memory.

Put in: exact test, typecheck, and lint commands; places where the code is misleading; decisions that live outside the repo; anything you've corrected twice. Leave out your folder layout, because Claude reads the repo itself. This template follows Anthropic's own example:

Settings are strict JSON (no comments, no trailing commas). Highest priority first:

1. Managed (managed-settings.json): set by your organization.
2. Command line: --settings <file-or-json> for one session.
3. Project local: .claude/settings.local.json, just you, kept out of git.
4. Shared project: .claude/settings.json, committed for the team.
5. User: ~/.claude/settings.json, you, every project.

Other locations: subagents in ~/.claude/agents/ and .claude/agents/; skills in ~/.claude/skills/<name>/SKILL.md and .claude/skills/<name>/SKILL.md; MCP servers in ~/.claude.json and .mcp.json at the project root. Run /status to confirm a file loaded. This example lets lint and tests run without asking and blocks reading .env files:

A subagent is a helper with its own context window, instructions, and tool list. Claude hands it a task and gets a summary back. That's useful for parallel research, for keeping noisy output out of your chat, and for review by fresh eyes. Built-in ones include Explore (fast, read-only search) and Plan.

To make one, ask Claude ("create a read-only code-reviewer subagent in .claude/agents/ that uses Sonnet") or write the file yourself. To use it, name it in your prompt, or @-mention it to guarantee it runs. Anthropic's own caveat: delegating doesn't make an answer better by itself, and each subagent costs extra tokens.

A skill is a saved procedure: a folder with a SKILL.md file. The folder name becomes a slash command, and Claude can also load it on its own when your request matches its description. Until then only the description sits in context, so long playbooks cost almost nothing. Custom commands were merged into skills: .claude/commands/deploy.md still works and creates /deploy, just like .claude/skills/deploy/SKILL.md.

Set disable-model-invocation: true for anything with side effects, like deploys, so only you can start it. Set user-invocable: false for background knowledge only Claude should load. Skills follow the open Agent Skills standard. This one is adapted from Anthropic's best-practices page:

A hook is a shell command that Claude Code (the program, not the model) runs automatically at a fixed point, so it fires every time. Anthropic's rule of thumb: if you keep adding emphasis to a CLAUDE.md line and Claude still misses it, that line belongs in a hook.

The events you'll use most: PreToolUse (before a tool runs, can block it), PostToolUse (after a tool succeeds, good for formatting), Stop (when Claude tries to finish; exit code 2 sends it back to work), and Notification (when Claude needs your input). There are more than 25 others. Hooks go in a hooks block in any settings file; check them with /hooks. This example from the docs runs Prettier on every file Claude edits:

MCP (Model Context Protocol) is an open standard for plugging tools into AI apps: a browser, Notion, Sentry, your database. If a system already has a CLI (like gh for GitHub), Claude can usually just use that instead.

• Remote servers use --transport http and a URL. For local servers, everything after -- is the command that starts them.
• The default scope is local (you, this project). --scope user covers all your projects; --scope project writes .mcp.json for the team.
• If a server needs a login, finish it from /mcp inside a session.
• Every connected server adds tool descriptions to your context, so turn off the ones you don't use.

Want to write your own? See how to build an MCP server.

claude -p "prompt" (--print) runs one task and exits. That's how Claude Code runs in CI jobs, git hooks, and shell loops. Useful flags:

• --output-format json returns one result object; stream-json with --verbose streams events.
• --allowedTools pre-approves tools, and --permission-mode dontAsk refuses everything else.
• --max-turns and --max-budget-usd cap how far a run can go.
• --bare skips hooks, skills, plugins, MCP servers, and CLAUDE.md so runs behave the same everywhere. It needs ANTHROPIC_API_KEY or cloud credentials, not your claude.ai login.
• --continue or --resume <id> carries a conversation across several -p calls.

Anthropic's advice: try a loop like this on three files first, fix the prompt, then run the rest.

These come from Anthropic's own best-practices page. Most bad sessions come down to one of four things: Claude was missing information, had no way to check its work, broke a rule nothing enforced, or spent more than the task was worth.

1. Tell Claude what it can't see. Name the file, describe the symptom, say what "fixed" looks like. Not "fix the login bug", but "users report login fails after the session times out. look at token refresh in src/auth/, write a failing test that reproduces it, then fix it."
2. Hand over the evidence. @ the file, paste the screenshot, pipe in the log.
3. Say "be ambitious" when you want the full version. Claude reads requests conservatively and builds the smallest thing that works. Only you know the scope you want.
4. Not sure what you want? Ask to be interviewed. Let Claude ask questions until it can write SPEC.md, then build from it in a fresh session.
5. Give it a check that passes or fails, like tests or a build, and ask for the command and its output. "All tests pass" on its own tells you little.
6. Review with fresh eyes. Use /code-review, or a subagent that checks against your spec. Add "skip style", or the reviewer will always find something.
7. Steer early. Watch the first minute and press Esc when Claude drifts. After two failed corrections, /clear and write one better prompt.
8. /clear between unrelated tasks. Every turn re-sends the whole conversation, so old tasks keep costing you.

What you can stop doing: ALL CAPS and "IMPORTANT" (use a hook instead), "think harder" (that's what /effort is for, though ultrathink still asks for deeper reasoning on one turn), and "don't be lazy" (give it a check to pass).

• New to Claude Code? Follow the Claude Code tutorial and build a small app end to end, then pick a project from 10 Claude Code projects with prompts.
• Weighing other tools? See Claude Code alternatives compared.
• Want structured learning? See the best Claude Code courses and the best vibe coding courses.
• Want to build agents, not just use one? Claude Code is a model in a loop with tools, memory, and permissions. Our free AI Agents with PydanticAI course teaches you to build those pieces in Python, starting with your first agent and tools.