Claude Code Cheat Sheet — Every Command, Shortcut, and Hidden Trick (2026)

If you bookmark only one block on this page, make it this one:

Action	How
Initialize project (create CLAUDE.md)	/init
Compact conversation context	/compact
Clear current session	/clear
List custom slash commands	/help
Switch model	/model
List active MCP servers	/mcp
Edit CLAUDE.md from inside session	/memory
Run a shell command without leaving Claude	!ls -la
Reference a file in the prompt	@src/auth.ts
Add a fact to project memory	# API rate limit is 60/min
Switch to plan mode	Shift+Tab twice
Cancel current operation	Esc
Edit a previous message	Esc once
Jump to any earlier message	Esc twice
Exit Claude Code	/exit or Ctrl+D

The rest of this guide covers the long tail. For deep dives on specific subsystems see our .claude/ folder guide, hooks guide, and Top 10 MCP servers ranking.

Type / in any session to get autocomplete for all of these. Versions ship updates frequently; if a command is missing in yours, run /help to see your version's exact list.

Core Workflow

Command	Effect
/help	List all available commands (built-in + custom)
/clear	Clear current session, start fresh
/compact [instruction]	Summarize old context to free token budget. Optional instruction guides what to keep.
/init	Generate a starter CLAUDE.md by scanning the project
/memory	Open CLAUDE.md (or .claude/CLAUDE.md) in your editor
/exit	Quit the session

Configuration

Command	Effect
/config	Open the settings UI (theme, model, keybindings)
/model	Switch model mid-session (Opus / Sonnet / Haiku)
/permissions	Edit allow/deny lists for tools
/agents	List, edit, or create custom subagents
/mcp	List active MCP server connections + their tools
/hooks	Inspect configured hooks for the current session
/login / /logout	Switch Anthropic accounts or re-auth

Session Management

Command	Effect
/resume	Pick a previous session to continue
/cost	Show token usage and estimated cost for current session
/status	Show current model, mode, working directory
/doctor	Diagnostics for installation issues

Most are universal across CLI and IDE extensions. Mac uses Cmd where Ctrl is shown; system shortcuts may override on Windows/Linux.

Shortcut	Effect
Esc	Cancel current operation; if input is empty, jump to previous message to edit
Esc Esc (double-tap)	Open message picker — jump to any earlier message in the conversation
Tab	Autocomplete file path, command name, or option
Shift+Tab	Cycle through modes: default → plan → auto-accept → default
↑ / ↓	Cycle through prompt history
Ctrl+R	Toggle verbose output (show full tool input/output)
Ctrl+C	Cancel currently-running tool; press twice to exit session
Ctrl+D	Exit session (when input is empty)
Ctrl+L	Clear screen (terminal only; doesn't clear conversation)
Ctrl+J	Insert literal newline (vs Enter which submits)
Alt+Enter / Opt+Enter	Same as Ctrl+J — newline without submit

Modes explained:

• Default: Claude asks before each tool call that isn't already allowed
• Plan: Claude only reads, never edits. Useful for "let's discuss the approach first"
• Auto-accept: Claude executes everything in permissions.allow without prompting. Use carefully on production code

Three single-character prefixes change how Claude treats your input. Mostly missed by new users; deeply useful once you know them.

! — Run a shell command

Anything after ! at the start of a prompt runs as a shell command, with the output included in the conversation:

!git log --oneline -5
!find . -name "*.test.ts" | headCOPY

The output gets summarized into the conversation context, which Claude can then reason over. Saves the "let me run a command, paste the output, then ask Claude" dance.

# — Add to project memory

Anything after # gets appended to CLAUDE.md as a remembered fact:

# API rate limit is 60 requests per minute
# Production database is read-replica only, writes go to primary at db-writer.internalCOPY

Useful for capturing facts in the moment instead of remembering to update CLAUDE.md later.

@ — Reference a file or symbol

@ with autocomplete points Claude at a specific file or symbol:

Refactor @src/auth/login.ts to use the new session API.
What does @validateUser do, and where is it called?COPY

Claude reads the referenced file before responding. Faster than asking "look at src/auth/login.ts" because the path is verified at autocomplete time.

Run claude --help for the canonical list. The flags worth memorizing:

Flag	Effect
claude	Start interactive session in current directory
claude -p "<prompt>"	Print mode: run prompt non-interactively, output to stdout, exit. Pipe-friendly.
claude --resume	Pick a previous session from a list
claude --continue	Continue the most recent session in this directory
claude --add-dir <path>	Add another directory to Claude's working set (without changing cwd)
claude --model <name>	Override default model (e.g. claude-opus-4-5)
claude --permission-mode <mode>	Set permission mode at startup: default, plan, acceptEdits, bypassPermissions
claude --output-format json	(Print mode) emit JSON instead of plain text. Good for scripting.
claude --include-partial-messages	(Print mode) include intermediate tool calls in output
claude mcp add <name> -- <cmd>	Register a new MCP server
claude mcp list / get / remove	Manage MCP servers without an editor
claude doctor	Diagnose installation, IDE integration, and connectivity
claude config get/set/list	Manage settings.json from CLI

Print mode pipeline trick

cat error.log | claude -p "summarize the most common error patterns" --output-format json | jqCOPY

Pipe-friendly print mode turns Claude into a proper Unix tool. Combine with find, grep, jq, awk for serious power.

Where everything lives. Hierarchy goes from most-global to most-specific; later files override earlier:

Path	Purpose
~/.claude/settings.json	Global user settings (apply to all projects)
~/.claude/CLAUDE.md	Global standing instructions
~/.claude.json	Account state, MCP server registrations, oauth tokens
~/.claude/agents/	Global subagents (available to all projects)
~/.claude/commands/	Global slash commands
./CLAUDE.md	Project standing instructions (committed to Git)
./.claude/settings.json	Project settings (committed)
./.claude/settings.local.json	Personal overrides (gitignored)
./.claude/agents/	Project subagents (committed)
./.claude/commands/	Project slash commands (committed)
./.mcp.json	Project-scoped MCP servers (committed via --scope project)

Eight tricks that separate confident users from beginners.

1. Compact with intent

/compact alone summarizes generically. /compact keep the file paths and the failed test names tells the summarizer what matters. Massive context recovery for long sessions.

2. Use Esc-Esc to fork the conversation

Realized 20 messages ago you should have asked differently? Esc Esc opens a message picker; pick the message, edit it, and Claude continues from that branch. The original branch is preserved.

3. Symlink CLAUDE.md to share across siblings

If you have multiple repos that should share standing instructions:

ln -s /shared/CLAUDE.md ~/work/repo-a/CLAUDE.md
ln -s /shared/CLAUDE.md ~/work/repo-b/CLAUDE.mdCOPY

Edit once, applies everywhere.

4. Pipe Claude's output to another command

Print mode + --output-format json + jq = serious automation:

claude -p "list all TODO comments in src/" --output-format json | jq -r '.text'COPY

5. Background long tasks with run_in_background

Inside a session, ask Claude to run a long-running task with run_in_background. Useful for builds and tests that would otherwise block the conversation.

6. Use --add-dir for monorepo workflows

If you need Claude to also see ~/code/shared-lib while working in ~/code/main-app, launch with claude --add-dir ~/code/shared-lib. Or run /add-dir mid-session.

7. Status line for git branch

Add to settings.json:

{
  "statusLine": {
    "type": "command",
    "command": "echo $(git branch --show-current 2>/dev/null) | tr -d '\n'"
  }
}COPY

Now the bottom of Claude Code always shows your branch. Catches "I'm committing on main" before it happens.

8. Keep a global ~/.claude/CLAUDE.md for personal preferences

Things like "be concise", "don't add comments unless I ask", "use TypeScript strict mode" belong in your global CLAUDE.md. Project CLAUDE.md adds project-specific stuff on top.

How do I see which keybindings are bound?

Run /config and navigate to the keybindings section. You can also check ~/.claude/keybindings.json directly.

What's the fastest way to undo a wrong edit Claude made?

Three options: (1) git checkout -- <file> if the change isn't committed; (2) git reset HEAD~1 if it's committed but not pushed; (3) Esc Esc to fork the conversation before the bad edit and try a different approach.

Can I run Claude Code in a sandbox?

Yes. Use --permission-mode plan for strict read-only, or run inside a Docker container with limited filesystem access. Claude Code respects permissions.deny patterns even with broad allow lists.

How do I share a slash command with someone outside my repo?

The slash command file is just Markdown. Send them .claude/commands/foo.md and ask them to drop it in their ~/.claude/commands/ (global) or their project's .claude/commands/.

Why doesn't /init generate everything I need in CLAUDE.md?

It generates a starter based on what it finds. Treat the output as a draft — review and trim. The good CLAUDE.md is what you keep after deleting half of what /init wrote.

How do I use Claude Code in CI without an interactive prompt?

Print mode: claude -p "<prompt>". Combined with --output-format json and --max-turns N, it's scriptable. Set ANTHROPIC_API_KEY in your CI environment.

Where do I report a bug or request a feature?

github.com/anthropics/claude-code/issues is the canonical issue tracker. Search before filing — many issues are duplicates.

Related: .claude/ folder complete guide · Hooks practical guide · Top 10 MCP servers