The Claude Code Cheat Sheet

Claude Code now spans the terminal, VS Code, JetBrains, desktop, browser, Slack, and cloud sessions. That makes old one-surface cheat sheets feel wrong fast.

This is the version I actually want handy now: install, login, permission modes, sandboxing, model picks, commands, and the workflow patterns that still hold up on April 13, 2026.

---

Start Here

Two current facts worth locking in:

• Claude Code is no longer just a terminal + API key tool. Official docs now support Claude subscriptions, Anthropic Console logins, and Bedrock, Vertex, or Foundry auth.
• Permission modes are now a first-class concept. Think in terms of default, acceptEdits, plan, auto, dontAsk, and bypassPermissions, not the older “just skip permissions” mental model.

---

Install and Launch

The official install path now starts with the native installer, not npm install -g.

# macOS, Linux, WSL
curl -fsSL https://claude.ai/install.sh | bash

# Homebrew stable channel
brew install --cask claude-code

# Homebrew latest channel
brew install --cask claude-code@latest

# Windows PowerShell
irm https://claude.ai/install.ps1 | iex

# Start a session in your project
cd your-project
claude

On first launch Claude opens a browser login flow. The official docs say you can authenticate with:

If you still use ANTHROPIC_API_KEY, that still works in the terminal CLI. Important nuance: if the key is set, it can take precedence over your subscription login once approved.

---

CLI Basics

# Start an interactive session
claude

# Start interactive with an opening task
claude "Audit this repo for dead code and suggest deletions"

# Print a one-shot answer and exit
claude -p "Explain the auth flow in this project"

# Start in plan mode
claude --permission-mode plan "Design the migration before touching code"

# Unlock auto mode in the Shift+Tab cycle
claude --enable-auto-mode

# Start a remote web session
claude --remote "Fix the login bug"

# Pull a web session back into your terminal
claude --teleport

# Resume a prior session
claude --resume

Important distinction: claude "task" opens a normal interactive session. claude -p "task" is the better fit for scripts, piping, CI, and one-shot runs.

---

Supported Models

Claude Code lets you choose aliases or pin full model names. The aliases are the easiest way to stay current; pin exact names only when reproducibility matters more than freshness.

Model	What it points to now	Best for
sonnet	Latest Sonnet model, currently Sonnet 4.6	Daily coding tasks, normal implementation loops
opus	Latest Opus model, currently Opus 4.6	Deep analysis, architecture, migrations, hard bugs
haiku	Latest Haiku model	Lightweight tasks and simpler asks
opusplan	Uses opus in plan mode, then sonnet for execution	Big tasks where planning quality matters more than execution cost
sonnet[1m]	Sonnet with 1M-token context	Long sessions or giant repos
opus[1m]	Opus with 1M-token context	Expensive but useful when context is the bottleneck

Model aliases always move forward. If you need stability, pin the full model name instead of relying on sonnet or opus.

# Explicit alias
claude --model opus "Review this migration plan for failure modes"

# Switch during a session
/model sonnet

Effort Levels

Claude Code supports effort controls directly in-session.

/effort high
/effort auto

---

Permission Modes

This is the biggest conceptual shift in Claude Code right now.

The docs define six permission modes:

# Start in plan mode
claude --permission-mode plan

# Start in bypass mode
claude --permission-mode bypassPermissions

# Shortcut form
claude --dangerously-skip-permissions

During a session, Shift + Tab cycles permission modes. The docs say the default cycle is default -> acceptEdits -> plan, with optional auto and bypassPermissions added only if you explicitly enable them.

---

Sandboxing

Permission modes and sandboxing are separate.

• Permission mode decides when Claude asks.
• Sandbox mode decides what shell commands are allowed to do.

Current Claude Code docs expose sandboxing in supported environments and a /sandbox command to toggle it. If you need Bash commands constrained, use sandboxing plus permission rules instead of relying on prompts alone.

/sandbox
/permissions

The useful mental model: permissions are about approval policy, sandboxing is about blast radius.

---

Commands Worth Memorizing

---

The Flags That Matter

--model <alias|name>                Pick a model or pinned version
--permission-mode <mode>            default | acceptEdits | plan | auto | dontAsk | bypassPermissions
--enable-auto-mode                  Add auto mode to the Shift+Tab cycle
-p, --print                         One-shot output without interactive mode
--output-format <format>            text | json | stream-json
--resume                            Resume a session
--remote                            Start a web session on claude.ai/code
--remote-control                    Let the same session be controlled from claude.ai
--teleport                          Pull a web session back into your terminal
--plugin-dir <path>                 Load plugins from a directory for this session
--settings <json-or-path>           Load extra settings
--dangerously-skip-permissions      Start in bypassPermissions mode

The two launch-time flags I’d keep on the tip of my tongue are --permission-mode and --model.

---

CLAUDE.md, Skills, Plugins, Hooks, MCP

Minimal CLAUDE.md I’d Actually Start With

# Project

## Stack
- Astro 5
- TypeScript
- Tailwind 4

## Commands
- Dev: `pnpm dev`
- Build: `pnpm build`
- Test: `pnpm test`

## Rules
- Propose a plan before touching more than 3 files
- Run targeted tests after edits
- Never edit `.env` or deployment config without asking

MCP Today

Tool search is now enabled by default in Claude Code. That matters because MCP servers no longer have to dump every tool into context up front when the model supports deferred tool references.

/mcp

If you run through a non-first-party proxy, the docs say tool search may be disabled unless you set ENABLE_TOOL_SEARCH explicitly.

---

Prompt Patterns That Still Work

Don’t say “fix this”

Set the blast radius

Use plan mode on purpose

Tell Claude how to finish

---

Quick Reference

claude                               Start an interactive session
claude -p "task"                     One-shot print mode
claude --resume                      Resume a past session
claude --remote "task"               Start a web session
claude --teleport                    Pull web work back local
claude --permission-mode plan        Inspect and plan only
claude --enable-auto-mode            Add auto mode to the cycle

/plan                                Switch to plan mode
/permissions                         Manage allow/ask/deny rules
/sandbox                             Toggle sandboxing
/model                               Change model
/effort                              Change effort
/plugin                              Manage plugins
/mcp                                 Manage MCP connections
/memory                              Edit CLAUDE.md and auto memory
/diff                                Review changes
/tasks                               View background tasks
/schedule                            Cloud-scheduled tasks

---

Three Rules I’d Keep

1. Start with CLAUDE.md, not vibes.
2. Use plan or acceptEdits deliberately instead of jumping straight to bypass mode.
3. Review the diff every time, even when the agent feels “obviously right.”