Claude Code Cheatsheet

⚡

The Core Workflow

Master This

0. PRD
Define with Claude

→

1. EXPLORE
Understand the codebase

→

2. PLAN
Shift+Tab → strategy

→

3. CODE
Build with verification

→

4. COMMIT
Git commit + PR

→

5. VERIFY
Tests + Chrome + PRD

💡 #1 Tip — Verification Loops: Always give Claude a way to verify its own work = 2-3x better quality. "After you build, run tests. If they fail, fix and rerun until green."

⌨️

Keyboard Shortcuts

Official

Key	Action
Shift+Tab	Cycle modes: Normal → Auto-Accept → Plan
Opt+T	Toggle extended thinking new
Ctrl+S	Stash current prompt new
Esc Esc	Rewind — undo code + conversation
Ctrl+C	Cancel current generation
Ctrl+D	Exit Claude Code session
Ctrl+B	Push running task to background
Ctrl+R	Reverse-search prompt history
Ctrl+T	Toggle task list view
Ctrl+L	Clear terminal screen
Ctrl+O	Toggle verbose output
Ctrl+V	Paste image from clipboard
Shift+Enter	New line (native since v2.1)

VS Code: Cmd+Esc toggle focus · Cmd+Opt+K insert @file
Tip: Press ? to see shortcuts for your terminal

Built-in Slash Commands

SESSION & CONTEXT:

/clear	Reset context use between tasks!
/compact	Compress history (at ~70%)
/context	Visualize context usage grid
/resume	Resume a previous session
/rewind	Rewind conversation + code
/export	Export conversation to file

CONFIG & TOOLS:

/init	Generate CLAUDE.md for project
/model	Switch model mid-session
/memory	Edit CLAUDE.md files directly
/permissions	Manage tool permissions
/config	Open settings UI
/doctor	Health check installation

FEATURES:

/chrome	Connect Chrome for visual tests
/sandbox	Sandboxed bash (safer YOLO)
/review	Request code review
/security-review	Security audit of changes
/skills	List available skills
/mcp	Manage MCP server connections
/todos	List current task items
/cost	Show token usage stats
/usage	Plan limits & rate status
/hooks	Manage automation hooks
/plugin	Plugin marketplace
/agents	Manage custom subagents
/tasks	List & manage background tasks new

Quick Prefixes & CLI

IN-SESSION PREFIXES:

!	Run bash directly (no AI interpretation)
@	File path autocomplete & reference
/	Slash command

CLI FLAGS:

        claude                 # Interactive REPL

        claude "query"        # Start with prompt

        claude --continue     # Continue last session

        claude --resume       # Pick a session to resume

        claude -p "query"     # Headless (CI/scripts)

        claude --chrome       # With Chrome browser

        claude --model opus   # Specify model

        claude --add-dir ../lib # Add extra dirs

        claude --debug        # Debug mode

        claude --dangerously-skip-permissions

⚠️ Prefer /permissions or /sandbox over --dangerously-skip-permissions

INSTALL:

        # Mac/Linux:

        curl -sL https://install.anthropic.com | sh

        # Windows:

        irm https://console.anthropic.com/install.ps1 | iex

        # Or via npm:

        npm install -g @anthropic-ai/claude-code

        # Update:

        claude update

📋

Plan Mode & Thinking

Key Skill

Start almost every non-trivial task in Plan Mode.

Shift+Tab cycles: Normal → Auto-Accept → Plan Mode → Normal

        # The flow:

        1. Enter Plan Mode (Shift+Tab ×2)

        2. Describe the task

        3. Read & refine Claude's plan

        4. Approve → switch to Auto-Accept

        5. Claude builds autonomously

        6. Verify results

A good plan includes: what to build, files affected, risks, and how to verify.

💡 Optional: Copy-paste the plan to your Claude project (claude.ai) and let Claude verify it before approving.

EXTENDED THINKING:

Auto-enabled by default. Toggle with Opt+T. Turn off for simple tasks to save tokens. Best for: architecture decisions, complex bugs, choosing between approaches.

✅

Verification Loops

Most Important

Don't trust — measure. Always give Claude a way to check its own work.

Method	How
Tests	`pytest -v` / `npm test`
Types	`tsc --noEmit`
Lint	`eslint .` / `ruff check`
Build	`npm run build`
Chrome	Visual check with extension
PRD	"Where are we vs the PRD?"

        # Template prompt with verification:

        "Build the feature. After you finish:

         1. Run the tests

         2. Open in Chrome and verify UI

         3. If anything fails → fix and rerun

         4. Repeat until all green"

🧠

Context Management

Claude has 200K token context. Manage it or quality drops.

<50%	All good, keep working
50-70%	Consider `/compact`
>70%	`/compact` now or `/clear`

Rule: Finished a task? /clear. Starting something new? /clear. Claude hallucinating? Check /context first.

@-MENTIONS:

@path/to/file.ts — include specific file
@docs/api-ref.md — include documentation
⚠️ Don't @-mention files inside CLAUDE.md — it bloats context permanently. Tell Claude to read them on demand instead.

🚀 Project Kickoff & Development Loop

🗺️

Project Kickoff Workflow

Follow This

1	Create project folder + Claude project — create a folder and a Claude project on claude.ai with relevant docs.
2	Build PRD with Claude — describe your project in conversation. Claude asks questions, you answer. Save as `PRD.md` in your project folder.
3	Create a dedicated GitHub repo — `git init` + create remote repo. Add `.gitignore` and `.env.example`.
4	Prepare initial Claude Code prompt — in the same Claude chat that created the PRD, ask: "Write me a detailed first prompt for Claude Code." Claude already has full context.
5	Open VS Code → File → Open Folder — open your project folder in VS Code.
6	Open terminal → run `claude` — launch terminal, type `claude` (or `--dangerously-skip-permissions`).
7	Enter Plan Mode → paste prompt — `Shift+Tab` ×2 to enter Plan Mode. Paste the initial prompt you crafted in step 4.
8	Review plan carefully — 💡 Copy plan to your Claude project for a second opinion. Iterate until solid.
9	After implementation → `/init` — generate CLAUDE.md to capture project context for future sessions.

⚠️ PRD first, always. Jumping straight to Claude Code without a PRD = Claude builds something, but probably not what you wanted.

🔄

During Development (Repeating Loop)

Core Loop

1	Plan — Plan Mode (`Shift+Tab`) for every significant step
2	Consult — discuss with Claude (desktop/web) in parallel
3	Verify — `/verify-yourself`
4	Commit — commit and push
5	Update memory — `/update-claude-md`
6	Security check — `/security-review`

CONTEXT MANAGEMENT:

/clear between tasks · /compact mid-task · /clear on direction change

WHEN YOU'RE STUCK:

Use a research sub-agent: "Start an agent to research…" · Explore agent for codebase understanding

SPECIALIZED TASKS:

frontend-design skill for UI · Put API docs in docs/ folder for integrations

💡

Prompting Rules

1. Be specific
File, function, expected vs actual.

2. Give context
Use @file refs, show patterns.

3. Break it down
One task at a time. Big = bad.

BAD:

        "Fix the bug"

        "Build me an e-commerce site"

GOOD:

        "In @auth.py, login() returns

         None instead of raising on

         bad password"

        "Add logging matching the style

         in @src/services/user_service.py"

In CLAUDE.md: use "IMPORTANT", "YOU MUST", "NEVER" — emphasis words get more attention.

Claude Code Cheatsheet — Page 2: Configuration, Tools & Advanced

by Omri Tal · Botique AI Solutions · v2.5

📄

CLAUDE.md — Project Memory

GLOBAL — ~/.claude/CLAUDE.md

Create once manually. Applies to ALL projects.

            # Personal preferences

            ## Communication

            - Respond in Hebrew

            - Be direct, no fluff

            ## Workflow

            - IMPORTANT: Plan Mode for complex tasks

            - YOU MUST verify before done

            ## Never Do

            - Never commit .env files

            - Never force push without asking

PROJECT — ./CLAUDE.md

Create with /init. Commit to git & share with team!

            # Auto-generated + your additions

            ## Overview

            FastAPI backend, React frontend

            ## Commands

            - Run: `uvicorn main:app --reload`

            - Test: `pytest -v`

            ## Common Mistakes

            - Don't use raw SQL, use ORM

            - All dates in UTC

            - API responses need request_id

Key habit: Every time Claude makes a mistake → add it to Common Mistakes. This compounds into a smarter AI partner over time.

Load order: Global → Project (project overrides) · Edit: /memory mid-conversation · Emphasis: "IMPORTANT", "YOU MUST", "NEVER"

🤖

Models & Accounts

Model	Best for
Opus 4.5	Complex, architecture Strongest
Sonnet 4.5	Daily work, balanced Default
Haiku 4.5	Search, simple tasks Fast

Switch: /model or claude --model opus

Plan	Price	Access
Pro	$20/mo	All models
Max 5x	$100/mo	More usage
Max 20x	$200/mo	Heavy usage

Check limits: /cost (tokens) · /usage (rate limits)

🔀

Subagents & Tasks

BUILT-IN AGENTS:

Explore	Search & read only (Haiku)
Plan	Planning mode (Sonnet)
Task	General-purpose sub-agent

        # Research in background

        "Start an agent to research the

         auth patterns in our codebase"

        → Ctrl+B to push to background

        → Continue working with main agent

        → Agent notifies when done

        → /tasks to see running tasks

TASK LIST (NEW):

Claude auto-creates task lists for complex work. Ctrl+T to toggle. Tasks persist across /compact. Share across sessions with CLAUDE_CODE_TASK_LIST_ID=name claude

CUSTOM AGENTS:

Create in .claude/agents/name.md with YAML frontmatter (name, description, model, tools). ⚠️ No nesting — subagents can't spawn subagents.

🪝

Hooks — Automation

Code that runs automatically before/after Claude's actions. Configure in .claude/settings.json

PreToolUse Before · PostToolUse After · Notification Alerts

        // Auto-format after every write:

        {

          "hooks": {

            "PostToolUse": [{

              "matcher": "Write|Edit",

              "hooks": [{

                "type": "command",

                "command": "bun run format || true"

              }]

            }]

          }

        }

Ideas: auto-format · auto-lint · run tests on write · notify on complete
Debug: claude --debug to see hook execution

🧩

Skills & Plugins

Skills = instruction folders loaded on-demand (Claude decides when). Plugins = packaged bundles. Since v2.1.3, commands & skills are unified.

        # Plugin marketplace

        /plugin → discover → install

        # Anthropic official skills

        /plugin marketplace add anthropics/skills

        /plugin install example-skills@anthropics/skills

        # Community skills

        → www.skills.sh

RECOMMENDED:

frontend-design	Anti-"AI slop" UI design
code-simplifier	Clean up after sessions
document-skills	docx, pdf, xlsx, pptx
superpowers	Battle-tested workflows

Custom skills: .claude/skills/name/SKILL.md

🔌

MCP Servers

Connect Claude to external tools via ~/.claude/mcp.json

        {

          "mcpServers": {

            "github": {

              "command": "npx",

              "args": ["-y",

                "@modelcontextprotocol/server-github"],

              "env": { "GITHUB_TOKEN": "ghp_..." }

            }

          }

        }

server-github	Repos, Issues, PRs
server-postgres	Query DB directly
server-slack	Read/write Slack
server-filesystem	Remote file access
server-puppeteer	Browser automation

Manage: /mcp · MCP commands appear in /help

🌐

Chrome & Git Integration

CHROME:

claude --chrome # or /chrome in session

Claude can navigate, click, fill forms, screenshot, verify UI. Requires Chrome Extension v1.0.36+, Claude Code v2.0.73+, Pro/Max.
⚠️ Stops at logins/CAPTCHAs. Not supported: Brave, Arc, WSL.

GIT:

Claude handles git natively — commits, pushes, creates PRs.

        # Claude can do all of this:

        "Commit changes with clear message"

        "Push to main and open a PR"

        "Show git log for last 5 commits"

        # GitHub Actions integration:

        /install-github-app

        # Then tag @claude in PRs & Issues!

⚙️

Custom Slash Commands

.claude/commands/name.md (project) or ~/.claude/commands/name.md (personal)

        ---

        description: Verify codebase vs PRD

        allowed-tools: Bash(*), Read(*), Grep(*)

        argument-hint: [focus-area]

        ---

        ## Context

        !`git diff --stat`

        ## Task

        Analyze codebase against PRD.md.

        Focus on: $ARGUMENTS

        Report: done, in-progress, todo.

        Use subagents for thorough review.

Syntax: !`cmd` = shell preprocessing (runs before prompt) · @file = include file · $ARGUMENTS = all args · $1 $2 = positional args

Ideas: /commit-push-pr · /test-and-fix · /code-vs-prd · /backend-audit · /security-scan

File Structure Reference

      # Global (your machine) — one-time setup                 # Project (per repo — commit to git!)

      ~/.claude/                                          project/

      ├── CLAUDE.md       # your preferences            ├── CLAUDE.md          # project config (/init)

      ├── mcp.json       # MCP servers                  ├── .claude/

      ├── settings.json  # global permissions            │   ├── commands/      # slash commands

      ├── commands/      # personal commands             │   ├── agents/        # custom subagents

      └── tasks/         # persistent task lists          │   ├── skills/        # custom skills

                                                      │   └── settings.json  # project permissions

                                                      ├── docs/              # API docs (on demand)

                                                      └── .env.example       # env var template