Complete Claude Code Cheatsheet

Claude Code is Anthropic's official terminal coding agent — a peer of OpenAI Codex CLI, Aider, and Cursor — that reads/writes files in your project, runs shell commands, talks to Claude (currently Opus 4.6 / Sonnet 4.6 / Haiku 4.5), and supports MCP servers, custom slash commands, sub-agents, and hooks.

Installation
Authentication
Daily Use
Slash Commands
Keyboard Shortcuts
Headless / Non-interactive Mode
Permission Modes
CLAUDE.md (Project Memory)
Settings (settings.json)
Custom Slash Commands
Sub-agents
Hooks
MCP Servers
Plugins
IDE Integrations
GitHub Actions / CI
Models
Cost & Context Management
Troubleshooting
Quick Reference

Installation

Bash

1
2
3
4
5
6
7
8
9
10
11
12
# npm (cross-platform)
npm install -g @anthropic-ai/claude-code

# Homebrew
brew install --cask claude-code

# Direct binaries: https://github.com/anthropics/claude-code/releases

# Verify
claude --version
claude doctor          # diagnose env, auth, dependencies

Requires Node 18+ for the npm path. Works on macOS, Linux, and Windows (WSL2 recommended on Windows).

Authentication

Bash

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
# Anthropic API key (default)
export ANTHROPIC_API_KEY=sk-ant-...

# Or interactive OAuth (Claude Pro/Max plan subscribers)
claude login
claude logout

# Vertex AI
export CLAUDE_CODE_USE_VERTEX=1
export CLOUD_ML_REGION=us-east5
export ANTHROPIC_VERTEX_PROJECT_ID=my-gcp-project

# Amazon Bedrock
export CLAUDE_CODE_USE_BEDROCK=1
export AWS_REGION=us-west-2

# OpenAI-compatible / proxy endpoints
export ANTHROPIC_BASE_URL=https://my-proxy.example.com
export ANTHROPIC_AUTH_TOKEN=...

Daily Use

Bash

1
2
3
4
5
6
7
8
9
10
11
12
cd ~/projects/my-app
claude                                 # start REPL in current dir
claude "explain the auth flow"         # seed an initial prompt
claude --continue                      # resume the last session
claude --resume                        # interactive session picker
claude --resume <session-id>
claude --cd ~/other-repo               # work elsewhere
claude --model sonnet                  # pick model
claude --model opus
claude --model haiku
claude --add-dir ../shared             # extra dirs the agent may read/write

Inside the REPL

Type a prompt and hit Enter. Use:

@path to attach a file or directory
!command to run a shell command directly (no agent)
#text to append a memory note to CLAUDE.md
/command for slash commands (see below)

Slash Commands

text

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
/help                  show help
/init                  generate CLAUDE.md for this repo
/clear                 clear chat (frees context)
/compact [focus]       summarise older turns; keep focus topic
/cost                  show token + dollar usage so far
/model [name]          switch model in this session
/permissions           manage allow/deny rules
/config                edit settings
/login                 (re-)login
/logout
/agents                list / scaffold sub-agents
/hooks                 list / configure hooks
/mcp                   list / install MCP servers
/plugin                manage plugins
/ide                   connect to your IDE (VS Code / JetBrains)
/install-github-app    install the GitHub PR/Issues app
/release-notes         what's new in Claude Code
/bug                   file a bug
/doctor                run diagnostics
/exit                  quit

User-defined commands appear as /your-command once you've added them — see Custom Slash Commands.

Keyboard Shortcuts

text

1
2
3
4
5
6
7
8
9
10
11
12
Enter            submit
Shift+Enter      newline (or paste-newline auto-detected)
Esc              interrupt the current model call
Esc Esc          undo the last user message
Ctrl+C           cancel current step; press twice to quit
Ctrl+D           quit
Ctrl+L           clear screen (keeps history)
Ctrl+R           reverse-search history
Tab              autocomplete file paths, slash commands, agent names
Up / Down        history
Ctrl+J           pop-out the multi-line editor in $EDITOR

Headless / Non-interactive Mode

For scripting and CI:

Bash

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
# Single-shot prompt (prints final answer, exits)
claude -p "summarise CHANGELOG.md"
claude --print "lint and fix"

# Read prompt from stdin
git diff --cached | claude -p "Write a PR description with risk + tests."

# JSON output for tooling
claude -p "describe repo" --output-format json | jq .
claude -p "..." --output-format stream-json   # NDJSON streaming

# Constrain tools / dirs in CI
claude -p "Update CHANGELOG for $TAG" \
  --permission-mode acceptEdits \
  --allowed-tools "Read,Edit,Write,Bash(git diff:*)" \
  --add-dir docs

# Continue / resume in headless mode
claude -p --continue "now bump the version"
claude -p --resume <session-id> "..."

Exit codes: 0 success, non-zero on error.

Permission Modes

Claude Code asks before touching anything by default. Modes:

Mode	Behaviour
`default`	Ask before edits and shell.
`acceptEdits`	Auto-approve file edits; still ask for shell.
`plan`	Read-only "planning" mode — no edits, no shell. Useful first pass.
`bypassPermissions`	Auto-approve everything. Dangerous; use in sandboxes only.

Bash

1
2
3
4
claude --permission-mode plan
claude --permission-mode acceptEdits
claude --dangerously-skip-permissions          # = bypassPermissions

Allow/deny rules (persisted via /permissions):

text

1
2
3
Allow:  Edit, Write, Read, Bash(git status:*), Bash(npm test:*)
Deny:   Bash(rm:*), WebFetch

`CLAUDE.md` (Project Memory)

Claude Code reads memory files automatically:

text

1
2
3
4
./CLAUDE.md                 project-level
./CLAUDE.local.md           project-level, gitignored
~/.claude/CLAUDE.md         global (all projects)

Bootstrap one with /init. Keep it short and specific.

Markdown

1
2
3
4
5
6
7
8
9
10
11
12
# CLAUDE.md
- Use pnpm, not npm.
- All new code requires vitest tests next to source in __tests__/.
- Don't touch /vendor, /generated, or the migrations folder.
- Style: functional React, Tailwind, no class components.
- Commit format: Conventional Commits.

## Build & test
- `pnpm dev` to run locally
- `pnpm test --run` for non-watch tests
- `pnpm lint && pnpm typecheck` before commit

Append a memory mid-conversation with #:

text

1
2
> # always run `pnpm typecheck` after edits

Also respects .gitignore and an optional .claudeignore for paths Claude should not read.

Settings (`settings.json`)

text

1
2
3
4
~/.claude/settings.json         user
./.claude/settings.json         project (commit this)
./.claude/settings.local.json   project, gitignored

jsonc

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
{
  "model": "claude-sonnet-4-6",
  "permissions": {
    "allow": [
      "Edit",
      "Write",
      "Read",
      "Bash(git diff:*)",
      "Bash(pnpm test:*)",
      "Bash(pnpm lint:*)"
    ],
    "deny": [ "Bash(rm:*)", "WebFetch" ],
    "defaultMode": "acceptEdits"
  },
  "env": {
    "NODE_ENV": "development"
  },
  "hooks": {
    "PostToolUse": [
      { "matcher": "Edit|Write",
        "hooks": [{ "type": "command", "command": "pnpm exec prettier --write \"$CLAUDE_FILE_PATHS\"" }] }
    ]
  },
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "${env:GITHUB_TOKEN}" }
    }
  },
  "statusLine": {
    "type": "command",
    "command": "~/.claude/statusline.sh"
  }
}

Custom Slash Commands

Put markdown files in .claude/commands/ (project) or ~/.claude/commands/ (user). The filename becomes the command name.

Bash

1
2
# .claude/commands/release-notes.md

Markdown

1
2
3
4
5
6
7
8
---
description: Draft release notes from staged commits.
allowed-tools: ["Bash(git log:*)", "Bash(git diff:*)", "Read"]
argument-hint: <version>
---

Look at the staged commits and write release notes for version $ARGUMENTS.
Use Conventional Commits to group features / fixes / chores.

Then in the REPL:

text

1
2
/release-notes v1.4.0

Arguments are expanded via $ARGUMENTS (or $1, $2, …). Front-matter allowed-tools constrains what the model can call during that command.

Namespacing: subdirs become prefixes (.claude/commands/db/migrate.md → /db:migrate).

Sub-agents

Sub-agents are specialised personas with their own system prompt and tool allowlist. Store them in .claude/agents/<name>.md (project) or ~/.claude/agents/<name>.md (user).

Markdown

1
2
3
4
5
6
7
8
9
10
11
12
---
name: reviewer
description: Read-only code reviewer. Use after staging changes.
tools: ["Read", "Bash(git diff:*)", "Bash(git log:*)"]
model: opus
---

You are a senior staff engineer. Review the staged diff and produce:
1. A summary in 3 bullets.
2. Risk assessment.
3. Concrete suggestions with file:line refs.
Do not modify files.

Invoke:

text

1
2
> use the reviewer agent on the staged diff

Or programmatically:

Bash

1
2
claude -p --agent reviewer "Review the staged diff"

Manage with /agents.

Hooks

Hooks run shell commands on lifecycle events.

Event	Fires when
`PreToolUse`	Before a tool call (you can block it).
`PostToolUse`	After a tool call succeeds.
`UserPromptSubmit`	Each user message.
`SessionStart`	Once per session.
`SessionEnd`	When the session ends.
`Notification`	On notifications (e.g. needs attention).
`Stop`	When Claude finishes responding.
`SubagentStop`	When a sub-agent finishes.

jsonc

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
// .claude/settings.json
{
  "hooks": {
    "PostToolUse": [
      { "matcher": "Edit|Write",
        "hooks": [{ "type": "command", "command": "pnpm exec prettier --write \"$CLAUDE_FILE_PATHS\"" }] }
    ],
    "PreToolUse": [
      { "matcher": "Bash",
        "hooks": [{ "type": "command", "command": "~/.claude/guard-rm.sh" }] }
    ],
    "SessionEnd": [
      { "hooks": [{ "type": "command", "command": "git status -s | head" }] }
    ]
  }
}

A hook with a non-zero exit blocks the action; stdout is shown to Claude. Useful env vars in hooks: $CLAUDE_FILE_PATHS, $CLAUDE_TOOL_NAME, $CLAUDE_SESSION_ID.

MCP Servers

Claude Code is an MCP client. Two ways to register servers:

Per-project, in `.claude/settings.json`

jsonc

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/home/me/projects"]
    },
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..." }
    },
    "postgres": {
      "command": "uvx",
      "args": ["mcp-server-postgres", "postgresql://localhost/mydb"]
    }
  }
}

From the CLI

Bash

1
2
3
4
5
6
claude mcp add github npx -- -y @modelcontextprotocol/server-github
claude mcp add --scope user github npx -- -y @modelcontextprotocol/server-github
claude mcp list
claude mcp remove github
claude mcp restart github

In the REPL:

text

1
2
/mcp           list servers and tool inventories

Claude Code can also expose itself as an MCP server for other clients:

Bash

1
2
claude mcp serve

Plugins

Plugins are installable bundles of slash commands, sub-agents, hooks, and MCP servers.

Bash

1
2
3
4
5
claude plugin install <repo-url-or-marketplace-id>
claude plugin list
claude plugin update
claude plugin remove <id>

Or via the REPL: /plugin.

Plugin layout (a plugin is just a repo or folder):

text

1
2
3
4
5
6
7
8
9
10
11
my-plugin/
├── plugin.json           # name, description, version, entry points
├── commands/
│   └── deploy.md
├── agents/
│   └── reviewer.md
├── hooks/
│   └── settings.json
└── mcp/
    └── server.js

IDE Integrations

text

1
2
/ide                  # connect to a running IDE (VS Code / JetBrains)

Or install:

VS Code: marketplace extension "Claude Code".
JetBrains: plugin marketplace search "Claude Code".

IDE integration adds: diff view for proposed edits, click-to-open file refs, selection-as-context, and a side-panel REPL.

GitHub Actions / CI

Install the GitHub App from the REPL:

text

1
2
/install-github-app

Then in a workflow:

YAML

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
name: claude-review
on:
  pull_request:
    types: [opened, synchronize]

jobs:
  review:
    runs-on: ubuntu-latest
    permissions:
      contents: read
      pull-requests: write
    steps:
      - uses: actions/checkout@v4
      - uses: anthropics/claude-code-action@v1
        with:
          mode: review
          anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
          allowed_tools: "Read,Bash(git diff:*)"

Headless from any CI:

YAML

1
2
3
4
- run: npm i -g @anthropic-ai/claude-code
- env:
    ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
  run: |
    claude -p --permission-mode acceptEdits \
      --allowed-tools "Read,Edit,Write,Bash(git diff:*)" \
      "Generate release notes for ${{ github.ref_name }}"

Models

Bash

1
2
3
4
5
6
7
8
claude --model opus
claude --model sonnet
claude --model haiku
# Or pin an exact ID:
claude --model claude-opus-4-6
claude --model claude-sonnet-4-6
claude --model claude-haiku-4-5-20251001

Picks:

Opus 4.6 — deepest reasoning, biggest refactors.
Sonnet 4.6 — balanced default for most coding work.
Haiku 4.5 — fastest/cheapest; great for mechanical edits.

/model switches mid-session.

Cost & Context Management

text

1
2
3
4
/cost                  totals for this session (input/output tokens, $)
/compact <focus>       summarise older turns, keep focus topic in view
/clear                 reset chat history

Best practices:

Use --permission-mode plan for the first exploration pass.
Use /compact once a session gets long, instead of /clear, to keep continuity.
Add a .claudeignore to skip large generated dirs.
Reach for Haiku for mechanical edits; Opus when you need real design thinking.

Troubleshooting

Bash

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
claude doctor                          # diagnostics

# Auth
claude logout && claude login
unset ANTHROPIC_API_KEY                # if conflicting with OAuth

# Hanging request
# press Esc to interrupt the current model call

# Tool was denied unexpectedly
/permissions                           # inspect allow/deny rules

# MCP server not showing up
claude mcp list
claude mcp restart <name>
tail -f ~/.claude/logs/claude-code-*.log

# Sub-agent never invoked
/agents                                # is it listed? matching keyword in prompt?

# Reset everything
rm -rf ~/.claude
claude login

Logs: ~/.claude/logs/. Session transcripts: ~/.claude/sessions/.

Quick Reference

Bash

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
# Daily
claude
claude "do X"
claude --continue
claude --resume

# Headless
claude -p "do X"
claude -p --permission-mode acceptEdits "do X"
claude -p --output-format json "do X" | jq

# Project
/init                                  # create CLAUDE.md
$EDITOR CLAUDE.md
$EDITOR .claude/settings.json
$EDITOR .claude/commands/<name>.md
$EDITOR .claude/agents/<name>.md

# MCP / plugins
claude mcp add ...
claude plugin install ...

# Diagnostics
claude doctor
claude --version

Tip: keep a tight CLAUDE.md and a couple of project-scoped slash commands (e.g. /test, /release-notes). It pays for itself within a week.

Claude Code

Claude Code

Complete Claude Code Cheatsheet

Table of Contents

Installation

Authentication

Daily Use

Inside the REPL

Slash Commands

Keyboard Shortcuts

Headless / Non-interactive Mode

Permission Modes

`CLAUDE.md` (Project Memory)

Settings (`settings.json`)

Custom Slash Commands

Sub-agents

Hooks

MCP Servers

Per-project, in `.claude/settings.json`

From the CLI

Plugins

IDE Integrations

GitHub Actions / CI

Models

Cost & Context Management

Troubleshooting

Quick Reference

Continue Learning