claude-code-ultimate-guide/guide/cheatsheet.md

Claude Code Cheatsheet

1 printable page - Daily essentials for maximum productivity

Author: Florian BRUNIAUX | Founding Engineer @Méthode Aristote

Written with: Claude (Anthropic)

Version: 3.6.0 | Last Updated: January 2026

---

Essential Commands

Command	Action
/help	Contextual help
/clear	Reset conversation
/compact	Free up context
/status	Session state + context usage
/context	Detailed token breakdown
/plan	Enter Plan Mode (no changes)
/execute	Exit Plan Mode (apply changes)
/model	Switch model (sonnet/opus/opusplan)
/exit	Quit (or Ctrl+D)

---

Keyboard Shortcuts

Shortcut	Action
Shift+Tab	Cycle permission modes
Esc × 2	Rewind (undo)
Ctrl+C	Interrupt
Ctrl+R	Retry last operation
Ctrl+L	Clear screen (keeps context)
Tab	Autocomplete
Shift+Enter	New line
Ctrl+D	Exit

---

File References

@path/to/file.ts    → Reference a file
@agent-name         → Call an agent
!shell-command      → Run shell command

IDE	Shortcut
VS Code	Alt+K
JetBrains	Cmd+Option+K

---

Permission Modes

Mode	Editing	Execution
Default	Asks	Asks
Auto-accept	Auto	Asks
Plan Mode	❌	❌

Shift+Tab to switch modes

---

Memory & Settings (2 levels)

Level	macOS/Linux	Windows	Scope	Git
Project	.claude/	.claude\	Team	✅
Personal	~/.claude/	%USERPROFILE%\.claude\	You (all projects)	❌

Priority: Project overrides Personal

File	Where	Usage
CLAUDE.md	Project root	Team memory (instructions)
settings.json	.claude/	Team settings (hooks)
settings.local.json	.claude/	Your setting overrides
CLAUDE.md	~/.claude/ (Win: %USERPROFILE%\.claude\)	Personal memory

---

.claude/ Folder Structure

.claude/
├── CLAUDE.md           # Local memory (gitignored)
├── settings.json       # Hooks (committed)
├── settings.local.json # Permissions (not committed)
├── agents/             # Custom agents
├── commands/           # Slash commands
├── hooks/              # Event scripts
├── rules/              # Auto-loaded rules
└── skills/             # Knowledge modules

---

Typical Workflow

1. Start session      → claude
2. Check context      → /status
3. Plan Mode          → Shift+Tab × 2 (for complex tasks)
4. Describe task      → Clear, specific prompt
5. Review changes     → Always read the diff!
6. Accept/Reject      → y/n
7. Verify             → Run tests
8. Commit             → When task complete
9. /compact           → When context >70%

---

Context Management (CRITICAL)

Statusline

Model: Sonnet | Ctx: 89.5k | Cost: $2.11 | Ctx(u): 56.0%

Watch Ctx(u): → >70% = /compact, >85% = /clear

Enhanced statusline (ccstatusline): Add to ~/.claude/settings.json:

{ "statusLine": { "type": "command", "command": "npx -y ccstatusline@latest", "padding": 0 } }

Actions by Symptom

Sign	Action
Short responses	/compact
Frequent forgetting	/clear
>70% context	/compact
Task complete	/clear

Context Recovery Commands

Command	Usage
/compact	Summarize and free context
/clear	Fresh start
/rewind	Undo recent changes
claude -c	Resume last session (CLI flag)
claude -r <id>	Resume specific session (CLI flag)

---

Under the Hood (Quick Facts)

Concept	Key Point
Master Loop	Simple while(tool_call) — no DAGs, no classifiers
Tools	8 core: Bash, Read, Edit, Write, Grep, Glob, Task, TodoWrite
Context	~200K tokens, auto-compacts at 75-92%
Sub-agents	Isolated context, max depth=1
Philosophy	"Less scaffolding, more model" — trust Claude's reasoning

Deep dive: Architecture & Internals

---

Plan Mode & Think Levels

Feature	Activation	Usage
Plan Mode	Shift+Tab × 2 or /plan	Explore without modifying
OpusPlan	/model opusplan	Opus for planning, Sonnet for execution
Think	--think flag	Standard analysis (~4K tokens)
Think Hard	--think-hard flag	Deep analysis (~10K tokens)
Ultrathink	--ultrathink flag	Maximum depth (~32K tokens)

OpusPlan workflow: /model opusplan → Shift+Tab × 2 (plan with Opus) → Shift+Tab (execute with Sonnet)

Required for: features >3 files, architecture, complex debugging

---

MCP Servers

Server	Purpose
Serena	Indexation + session memory + symbol search
mgrep	Semantic search by intent (alternative)
Context7	Library documentation
Sequential	Structured reasoning
Playwright	Browser automation
Postgres	Database queries

Serena memory: write_memory() / read_memory() / list_memories()

Check status: /mcp

---

Creating Custom Components

Agent (.claude/agents/my-agent.md)

---
name: my-agent
description: Use when [trigger]
model: sonnet
tools: Read, Write, Edit, Bash
---
# Instructions here

Command (.claude/commands/my-command.md)

# Command Name
Instructions for what to do...
$ARGUMENTS - user provided args

Hook (macOS/Linux: .sh | Windows: .ps1)

Bash (macOS/Linux):

#!/bin/bash
INPUT=$(cat)
# Process JSON input
exit 0  # 0=continue, 2=block

PowerShell (Windows):

$input = [Console]::In.ReadToEnd() | ConvertFrom-Json
# Process JSON input
exit 0  # 0=continue, 2=block

---

Anti-patterns

❌ Don't	✅ Do
Vague prompts	Specify file + line
Accept without reading	Read every diff
Ignore warnings	Use /compact
Skip permissions	Never in production
Giant context loads	Load only what's needed
Negative constraints only	Provide alternatives

---

Quick Prompting Formula

WHAT: [Concrete deliverable]
WHERE: [File paths]
HOW: [Constraints, approach]
VERIFY: [Success criteria]

Example:

Add input validation to the login form.
WHERE: src/components/LoginForm.tsx
HOW: Use Zod schema, show inline errors
VERIFY: Empty email shows error, invalid format shows error

---

CLI Flags Quick Reference

Flag	Usage
-p "query"	Non-interactive mode
-c / --continue	Continue last session
-r / --resume <id>	Resume specific session
--headless	Non-interactive (CI/CD)
--model sonnet	Change model
--add-dir ../lib	Add directory
--permission-mode plan	Plan mode
--dangerously-skip-permissions	Auto-accept (use carefully)
--debug	Debug output
--mcp-debug	Debug MCP servers
--allowedTools "Edit,Read"	Whitelist tools

---

Debug Commands

claude --version     # Version
claude update        # Check/install updates
claude doctor        # Diagnostic
claude --debug       # Verbose mode
claude --mcp-debug   # Debug MCPs
/mcp                 # MCP status (inside Claude)

---

CI/CD Mode (Headless)

# Non-interactive execution
claude -p "analyze this file" src/api.ts

# JSON output
claude -p "review" --output-format json

# Economic model
claude -p "lint" --model haiku

# With auto-accept
claude -p "fix typos" --dangerously-skip-permissions

---

The Golden Rules

1. Always review diffs before accepting
2. Use /compact before context gets critical (>70%)
3. Be specific in requests (WHAT, WHERE, HOW, VERIFY)
4. Plan Mode first for complex/risky tasks
5. Create CLAUDE.md for every project
6. Commit frequently after each completed task
7. Know what's sent — prompts, files, MCP results → Anthropic (opt-out training)

---

Quick Decision Tree

Simple task       → Just ask Claude
Complex task      → TodoWrite to plan first
Risky change      → Plan Mode first
Repeating task    → Create agent or command
Context full      → /compact or /clear
Need docs         → Use Context7 MCP
Deep analysis     → Use --think or --ultrathink

---

Common Issues Quick Fix

Problem	Solution
"Command not found"	Check PATH, reinstall npm global
Context too high (>70%)	/compact immediately
Slow responses	/compact or /clear
MCP not working	claude mcp list, check config
Permission denied	Check settings.local.json
Hook blocking	Check hook exit code, review logic

Health Check Script (save & run):

# macOS/Linux
which claude && claude doctor && claude mcp list

# Windows PowerShell
where.exe claude; claude doctor; claude mcp list

---

Cost Optimization

Model	Use For	Cost
Haiku	Simple fixes, reviews	$
Sonnet	Most development
Opus	Architecture, complex bugs	$
OpusPlan	Plan (Opus) + Execute (Sonnet)

Tip: Use --add-dir to load only needed directories (saves tokens)

---

Resources

• Official docs: docs.anthropic.com/claude-code
• Advanced guide: Claudelog.com - Tips & patterns
• Full guide: ultimate-guide.md (this repo)
• Project memory: Create CLAUDE.md at project root
• DeepSeek (cost-effective): Configure via ANTHROPIC_BASE_URL

---

Author: Florian BRUNIAUX | @Méthode Aristote | Written with Claude

Last updated: January 2026 | Version 3.6.0