Claude Code Cheat Sheet

A quick reference for working with Claude Code and prompt engineering.

💡 Your system is detected automatically — Mac- and Windows-specific commands are filtered to it. Switch above if you need the other.

⌨️ Hotkeys

Mac-specific keys:

Windows-specific keys:

🖥️ CLI: launch and flags

claude                          # Interactive mode
claude "request"                # Launch with an initial request
claude -p 'request'             # Pipe mode: response to stdout and exit
claude -c                       # Resume the last session
claude -r <id>                  # Resume session by ID
claude update                   # Update Claude Code

# Useful flags
--model <model>                 # Specify model at launch
--permission-mode plan          # Launch in read-only mode
--worktree / -w                 # Isolated git worktree
--no-auto-compact / -n          # Disable auto context compaction
--add-dir <path>                # Add a folder to context
--output-format json|text|stream-json
--max-turns <N>                 # Turn limit in auto mode
--max-budget-usd <N>            # Session budget in dollars
--bare                          # Without CLAUDE.md and MCP (clean start)

# Pipe mode — handy for scripts:
cat error.log | claude -p 'explain the error'

📋 Session management

/clear                  # Full context wipe
/compact [topic]        # Compress context, keep the essentials
/context                # Show context fill
/resume                 # List recent sessions to resume
/rename <name>          # Rename the session (handy for /resume)
/branch                 # Create a session branch
/cost                   # Current session cost ($, tokens)
/diff                   # All file changes this session
/copy                   # Copy the last response
/export                 # Export the conversation to Markdown

Decision tree /clear vs /compact:

💡 CLAUDE.md is auto-reloaded after /compact and auto-compact — instructions aren't lost.

⚙️ Settings

/config                         # Open settings.json in the editor
/model <model>                  # Switch model in the session
/fast                           # Switch to Haiku (fast/cheap)
/effort low|medium|high         # Analysis depth (NEW)
/permissions                    # View and configure permissions
/vim                            # Enable vim editing mode
/theme                          # Change color theme

Permission modes:

🔧 Tools and extensions

/init               # Create CLAUDE.md for the project
/memory             # Open CLAUDE.md for editing
/mcp                # Manage MCP servers
/hooks              # Manage hooks
/skills             # Manage skills
/agents             # Manage agents
/chrome             # Browser tools via MCP
/reload-plugins     # Reload plugins without a restart

🧩 Slash commands

/btw <text>         # Add context without interrupting work
/plan               # Create a markdown task plan
/loop               # Iterative loop until a condition is met
/voice              # Voice input (Space = push-to-talk)
/doctor             # Diagnostics: API keys, MCP, configs
/rc                 # Open ~/.claude/rc.ts (launch settings)
/remote-control     # Control via API (NEW)
/pr-comments        # Load PR comments into context
/stickers           # 🥚

🧩 Plugins and skills

/plugin marketplace add <name>  # Install from the marketplace
/plugin list                    # List installed
/plugin enable/disable <name>   # Enable/disable a plugin
/simplify                       # Simplify code
/debug                          # Debug: error analysis
/batch                          # Bulk file changes

Skill locations:

• .claude/skills/ — project skills
• ~/.claude/skills/ — global skills
• .claude/agents/ — agents

Skill frontmatter:

---
description: "What the skill does"
allowed-tools: [Bash, Read, Write, mcp__github__*]
model: claude-sonnet-4-20250514
effort: medium          # low | medium | high (NEW)
context: fork           # isolated context
---
# Skill body
Task: $ARGUMENTS        # arguments are substituted here

🤖 Agent parameters

permissionMode: default|acceptEdits|plan|dontAsk|bypassPermissions
isolation: none|worktree    # worktree = separate git worktree
memory: path/to/MEMORY.md   # memory across runs
background: true            # background run
maxTurns: 10                # turn limit

Which agent when:

🌀 Ultrathink

Activate deep thinking — trigger words in the prompt:

think → think hard → think harder → ultrathink

Or the Alt+T key. Useful for: architectural decisions, tough bugs, ambiguous specs.

Pattern Plan → Normal → Auto:

1. Plan mode   → recon only, no changes
2. Normal mode → implementation with confirmation
3. Auto-accept → routine without confirmation (Shift+Tab)

🔄 A typical day with Claude Code

9:00  [You] Thinking about strategy
      └─→ 30 min of thinking, planning

9:30  [Agent] Makes the first draft
      └─→ Claude Code writes code/text

10:00 [You] Reviewing the result
      └─→ Review, feedback, clarifications

10:15 [Agent] Refines based on feedback
      └─→ Polished result

10:30 [You] Publish/deploy
      └─→ Final decision

🏃 Quick start (5 minutes)

# 1. Create and enter a project
mkdir my-project && cd my-project

# 2. Create context files
touch CLAUDE.md TODO.md AGENTS.md

# 3. Initialize CLAUDE.md
claude
/init       # ← inside Claude Code

# 4. Or fill in manually:
# Stack, how to run, agent instructions

# 5. Give it the first task

# 1. Create and enter a project
mkdir my-project; cd my-project

# 2. Create context files (PowerShell)
ni CLAUDE.md, TODO.md, AGENTS.md -ItemType File

# 3. Initialize CLAUDE.md
claude
/init       # ← inside Claude Code

# 4. Or fill in manually:
# Stack, how to run, agent instructions

# 5. Give it the first task

💡 Alternative via WSL/Git Bash: mkdir my-project && cd my-project && touch CLAUDE.md TODO.md AGENTS.md

📝 Four shifts (Quick Ref)

📋 Structure of a good prompt

## Role
You are — [description + expertise]

## Task
[Specific task]

## Input
[Input data]

## Output
[Expected result]

## Constraints
[Constraints]

🎯 Delegation matrix

✅ Delegate:

• Routine operations (formatting, conversion)
• First drafts (text, code)
• Information lookup
• Tests
• Documentation

🧠 Decide yourself:

• Architecture
• Final quality assessment
• Strategic decisions
• Ethical questions
• Ambiguous tasks

📄 Required project files

CLAUDE.md

Quick start:

/init  # Auto-generates CLAUDE.md from an existing project

Where to put it:

• ./CLAUDE.md — project (committed to git)
• ~/.claude/CLAUDE.md — personal (all projects)

Limit: ~200 lines (important for context)

Minimal structure:

# Project Context
## My role
## Project
## Stack
## How to run
## Instructions for Claude Code

TODO.md

# Tasks
## 🔴 CRITICAL
## 🟡 IMPORTANT
## 🟢 LATER
## ⏸️ BLOCKED
## ✅ DONE

AGENTS.md

# Agent Instructions
## For any work
## For /explore
## For /plan
## For /code
## What to avoid

⚠️ Common mistakes

💡 Quick tips

What to use when

For analyzing a codebase:

/explore --thoroughness medium

For planning a big piece of work:

/plan

For quickly adding info:

/btw Update the context: [info]

For exploring a specific file:

/explore --thoroughness quick
"Which files are responsible for [X]?"

🎯 Prompt quality check

Before sending:

• Role is clear?
• Task is specific?
• Data is complete?
• Output is defined?
• Constraints are clear?
• No distractors?
• Important stuff at the start/end?

🔢 Tokens

1 token ≈ 4 characters
1000 tokens ≈ 750 words

Cost:
- Input token = cheaper
- Output token = pricier (2–3×)

Optimize:
- Compress context
- Strip the extra
- Structure your data

Advanced token tracking

After Meeting 2 you can set up the advanced status line (see 02-setup-guide.md → "Advanced setup"):

• Multi-line display with daily and weekly budget
• Colored progress bars
• Tracking new sessions vs continuation
• Promotional discount support

🌍 Prompt examples

Backend request

## Role
You are a seasoned backend developer with 10 years of experience.
You know Node.js, PostgreSQL, microservices.

## Task
Create a REST API endpoint for fetching the user list.

## Input
- Fields needed: id, name, email, created_at
- Pagination support (page, limit)
- Filtering by email

## Output
- Express route handler
- TypeScript types
- Example curl requests

## Constraints
- Use a PostgreSQL query builder (knex.js)
- Max 50 lines of code
- Add validation for query parameters

Frontend request

## Role
You are a frontend developer, an accessibility specialist.

## Task
Create a React component for a product filter.

## Input
- Filters: category (checkbox), price (range), rating (radio)
- Sample data in data.json

## Output
- React component Filter.tsx
- Tailwind CSS styles
- WCAG AA accessibility

## Constraints
- TypeScript
- No UI libraries (pure CSS)
- Keyboard navigation (Tab, Enter, Escape)

Research request

## Role
You are a UX researcher experienced in user interviews.

## Task
Analyze interview.txt and find the main user pain points.

## Input
- File interview.txt with the interview recording

## Output
- Top 5 user problems
- For each: description, quote, possible solution

## Constraints
- Max 1 page
- Direct quotes from the interview only
- Prioritize by frequency of mentions

🔐 Git commands (quick)

Same on both systems — Git is identical everywhere:

git status                          # Status
git add .                           # Stage all files
git commit -m "Description"         # Commit
git push origin main                # Push
git pull origin main                # Pull

🖥️ Mac / Windows differences

Where the commands actually diverge (everything else is identical):

💡 Tip for Windows devs: install WSL2 + Git Bash so you can run bash commands from tutorials. PowerShell works too, but many AI blogs assume bash.

💡 Tip for Mac: use iTerm2 or Warp instead of the stock Terminal — they pair better with Claude Code (patterns, colors, hotkeys).

📊 Result check

# Result is good if:
- ✅ Code runs without fixes
- ✅ Structure is clear
- ✅ Usage examples included
- ✅ Edge cases covered
- ✅ Performance is acceptable

# Result is bad if:
- ❌ You have to finish by hand
- ❌ Details from the prompt are lost
- ❌ Hallucinations in the code
- ❌ Lots of unnecessary code

🎓 Going further

1. Run through all EXERCISES in EXERCISES.md
2. Create your own prompt template
3. Use it in real projects
4. Record good prompts
5. Share with your team

📞 Quick help

claude --help       # All CLI flags
/doctor             # Install diagnostics
/context            # Context fill
/cost               # Session cost
/btw                # Add context without interrupting
/plan               # Planning mode

Last updated: 2026-03-25

💡 Interactive version with 141 cards + quiz mode: tools/cheatsheet/index.html

Use this cheatsheet for a quick start and quick reference.