Claude Code Cheat Sheet

⌨ Keyboard Shortcuts

Essentials

EnterSend message / confirm EscCancel current input Esc EscRewind / summary menu Ctrl+CCancel input or generation Ctrl+DExit Claude Code Shift+TabCycle permission modes Ctrl+BSend task to background

Navigation & Display

Ctrl+RSearch history Ctrl+LClear screen (keeps conversation) Ctrl+OFullscreen transcript — cycles views NEW Ctrl+TShow/hide task list Ctrl+SStash current prompt ↑/↓Navigate command history ←/→Cycle tabs / footer

Editing & Model

\ + EnterMulti-line input Ctrl+GOpen in external editor Ctrl+VPaste image (screenshots, diagrams) Ctrl+KDelete to end of line Ctrl+UDelete from cursor to beginning Ctrl+YPaste deleted text Ctrl+_Undo last action Alt+B / Alt+FMove back / forward one word Option+PModel selector Option+TToggle thinking (adaptive on Opus 4.7) Alt+OToggle fast mode ?Show all shortcuts Space (hold)Push-to-talk (if /voice enabled) Alt+YCycle paste history (after Ctrl+Y) Ctrl+JNewline (control sequence) Option+EnterNewline (macOS) Ctrl+X Ctrl+KKill all background agents (×2 to confirm) Ctrl+X Ctrl+EOpen in external editor (alias Ctrl+G)

Vim mode NEW

vVisual mode — character selection VVisual-line mode — line selection EscBack to normal mode (improved)

Quick Prefixes

/Slash command or skill !Bash mode — run a shell command @Autocomplete file/folder/agent

Confirmation Dialogs

Y / EnterConfirm N / EscDeny Ctrl+EToggle permission explanation Ctrl+DToggle permission debug info

Tip: /terminal-setup for Shift+Enter. /keybindings to customize. Contexts: Global, Chat, Confirmation, Transcript, etc. (17 contexts). Flicker-free rendering (alt-screen) is the default since v2.1.89. Set CLAUDE_CODE_NO_FLICKER=0 to revert.

⚡ Slash Commands

Session & Context

`/clear`	Reset the conversation (alias: /reset, /new)
`/compact [focus]`	Compact with optional focus
`/resume [session]`	Resume by ID or name (alias: /continue)
`/rewind`	Go back to a checkpoint (alias: /checkpoint)
`/rename [name]`	Rename the session
`/branch [name]`	Branch the conversation (alias: /fork)
`/export [file]`	Export conversation as text
`/copy [N]`	Copy response N to clipboard
`/cost`	Costs + tokens + per-model & cache-hit breakdown (subscribers) NEW
`/context`	Color-coded context consumption grid
`/tasks`	List and manage background tasks

Model & Configuration

`/model [name]`	Switch model (sonnet, opus, haiku) — Opus 4.7 available NEW
`/effort [level]`	Interactive slider — low/medium/high/xhigh/max/auto NEW
`/fast [on\|off]`	Toggle fast output mode (Opus 4.6, 1M context)
`/focus`	Condensed view: last prompt + tool summary + response NEW
`/tui`	Flicker-free fullscreen rendering (alt-screen) NEW
`/config`	Config panel (includes Editor mode, replaces /vim) (alias: /settings)
`/permissions`	Manage perms + Recent tab (retry with `r`) NEW (alias: /allowed-tools)
`/mcp`	Manage MCP servers & OAuth
`/hooks`	View configured hooks
`/theme`	Change theme + custom JSON editor (~/.claude/themes/)
`/color [color]`	Prompt bar color
`/sandbox`	Toggle sandbox mode
`/vim`	Removed in v2.1.92 → `/config` → Editor mode REMOVED
`/tag`	Removed in v2.1.92 → AI-generated session titles REMOVED

Project & Analysis

`/init`	Create CLAUDE.md for the project
`/memory`	Edit CLAUDE.md & memory, toggle auto-memory
`/add-dir <path>`	Add a working directory
`/diff`	Interactive diff of changes
`/review`	Code review (deprecated → code-review plugin)
`/security-review`	Scan for vulnerabilities
`/insights`	HTML usage analysis report
`/stats`	Visualize daily usage, streaks, preferences
`/ultraplan`	Cloud drafting, browser review, local/remote exec NEW
`/pr-comments [PR]`	GitHub PR comments
`/autofix-pr`	Fix a PR from the terminal or web NEW
`/plan [desc]`	Enter plan mode with optional description
`/simplify`	Analyze code for reuse, quality, efficiency
`/batch [task]`	Parallel changes in worktrees
`/debug [desc]`	Session troubleshooting

Extensions & Plugins

`/plugin`	Manage plugins (auto-resolve deps) NEW
`/reload-plugins`	Reload skills/agents/MCP without restart
`/agents`	Tabs: Running (● N) + Library NEW
`/skills`	List sorted by token cost NEW
`/chrome`	Configure Claude in Chrome
`/ide`	Manage IDE integrations

Advanced & Remote

`/loop [interval]`	Repeat — omit interval for self-pacing NEW
`/schedule`	Cloud-scheduled tasks
`/btw`	Quick question without tools or history
`/voice`	Push-to-talk voice dictation
`/teleport`	Resume a web session locally
`/desktop`	Transfer to Desktop app
`/remote-control`	Make session accessible from claude.ai
`/remote-env`	Configure default remote environment

System & Account

`/help`	Help and command reference
`/doctor`	Check installation and auth
`/status`	Version and account info
`/usage`	Plan limits and rate limits
`/extra-usage`	Configure extra usage
`/feedback`	Submit a report (alias: /bug)
`/release-notes`	Interactive version picker NEW
`/powerup`	Animated lessons to learn Claude Code NEW
`/team-onboarding`	Generate a ramp-up guide for a teammate NEW
`/login / /logout`	Login / Logout
`/upgrade`	Upgrade to higher plan
`/privacy-settings`	Privacy settings
`/keybindings`	Open keybindings file
`/terminal-setup`	Configure terminal keybindings
`/statusline`	Configure status line + `refreshInterval` NEW
`/install-github-app`	Install the GitHub Actions app
`/install-slack-app`	Install the Slack app via OAuth
`/exit`	Exit (alias: /quit)
`/mobile`	QR code for mobile app (alias: /ios, /android)
`/passes`	Share a free week of Claude Code
`/stickers`	Order Claude Code stickers

🚀 CLI Launch Flags

Session Startup

`claude`	Interactive session
`claude "query"`	Start with an initial prompt
`claude -p "query"`	Print mode — output and exit
`claude -c`	Resume last conversation
`claude -r "id"`	Resume by name or ID
`claude -n "name"`	Name the session
`claude -w branch`	Isolated git worktree

Model & Reasoning

`--model sonnet`	Choose the model (sonnet/opus/haiku)
`--model opus[1m]`	Opus + 1M context (native on Opus 4.7)
`--model claude-opus-4-7`	Opus 4.7 — adaptive thinking NEW
`--effort xhigh`	Effort level xhigh (100K tokens) NEW
`--effort {low..max,auto}`	low / medium / high / xhigh / max / auto
`--fallback-model`	Fallback model if overloaded

Context & Directories

`--add-dir ../path`	Additional directories
`--bare`	Skip hooks, plugins, MCP, CLAUDE.md
`--verbose`	Detail for each turn
`--debug "api,hooks"`	Debug by categories
`--chrome / --no-chrome`	Enable/disable Chrome
`--ide`	Auto-connect IDE at startup

Permissions

`--permission-mode plan`	Read-only mode
`--permission-mode auto`	AI classifies permissions
`--allowedTools`	Auto-approved tools
`--disallowedTools`	Tools removed from context
`--tools`	Restrict built-in tools
`--enable-auto-mode`	Unlock Auto in Shift+Tab
`--permission-prompt-tool`	MCP tool for permissions (headless)
`--dangerously-skip-permissions`	Equivalent of bypassPermissions

Output & Formats

`--output-format json`	Structured JSON output
`--output-format stream-json`	Real-time streaming JSON
`--input-format`	Input format (text, stream-json)
`--include-partial-messages`	Include partial streaming events
`--json-schema '{}'`	Force an output schema
`--max-turns 5`	Limit agentic turns
`--max-budget-usd 2`	Spending limit

System Prompt

`--system-prompt "text"`	Override the system prompt
`--system-prompt-file`	Load from a file
`--append-system-prompt`	Append text to the prompt
`--append-system-prompt-file`	Append from a file

Advanced & Remote

`--remote "task"`	Remote session on claude.ai
`--remote-control`	Enable remote control
`--tmux`	Worktree + tmux pane
`--teammate-mode`	Agent teams: auto/in-process/tmux
`--agent name`	Run as agent
`--agents`	Define agents via JSON
`--from-pr 42`	PR resume: GitHub, GitLab, Bitbucket, GH Enterprise NEW
`--fork-session`	Fork while keeping context
`--teleport`	Resume a web session
`--mcp-config`	Load MCP servers
`--strict-mcp-config`	Use ONLY these MCP servers
`--plugin-dir`	Load plugins from a directory
`--settings`	Custom JSON settings file
`--init / --init-only`	Run init hooks (and exit)
`--maintenance`	Run maintenance hooks and exit
`--disable-slash-commands`	Disable skills and commands
`--no-session-persistence`	Disable persistence (print)
`--session-id`	Specific session UUID
`--agents`	Define dynamic agents via JSON
`--channels`	MCP notifications to listen to (research preview)
`--setting-sources`	Setting sources to load (user, project, local)
`--betas`	Beta headers for API requests

CLI Subcommands

`claude update`	Update
`claude auth login`	Login (--email, --sso, --console, --third-party) NEW
`claude auth status`	Auth status (JSON or --text)
`claude auth logout`	Sign out the account
`claude agents [list\|run]`	List and run agents
`claude mcp add/list/remove`	Manage MCP servers
`claude mcp add-from-claude-desktop`	Import from Desktop app
`claude mcp add-json`	Add from a JSON blob
`claude mcp serve`	Expose Claude Code as an MCP server
`claude plugin [install\|list\|marketplace]`	Manage plugins & marketplaces
`claude auto-mode defaults`	Default auto-mode rules
`claude auto-mode config`	Effective config with settings
`claude bedrock setup`	Interactive Bedrock wizard NEW
`claude vertex setup`	Interactive Vertex AI wizard NEW

🔌 The Big 5 — Claude Code Extension System

1. CLAUDE.md — System Prompt

`Managed`	Enterprise policy, highest priority
`Project`	`./CLAUDE.md` or `./.claude/CLAUDE.md`
`User`	`~/.claude/CLAUDE.md`

Imports: @docs/standards.md (max 5 levels)

Size: Keep < 200 lines, @path to split

2. Rules — Scoped Rules

Files in .claude/rules/*.md, activated by glob:

paths: ["src/**/*.tsx"]
MUST use TanStack Query.
SHOULD prefer named exports.

Placement based on consumer. MUST/SHOULD/MAY

3. Custom Commands — Shortcuts

Create in .claude/commands/name.md → /name

$ARGUMENTS, $0, $1 for arguments

Live context: !`git diff` (executed before prompt)

4. Skills — Complex Workflows

Directory .claude/skills/name/SKILL.md

`description`	Auto-discovery trigger (1536 chars) NEW
`context: fork`	Isolated sub-agent (no context pollution)
`paths: "*.ts"`	Activate by file type
`effort: high`	Force extended thinking
`user-invocable`	true = /command, false = Claude only
`allowed-tools`	Restrict tools used by the skill
`model`	Per-skill model override

Progressive disclosure: SKILL.md < 500 lines, [ref](references/). /skills sorts by token cost NEW.

Builtin: code-review, security-review, simplify, init, deslopify, loop, claude-api, session-start-hook.

5. Sub-agents — Parallel Specialists

Create in .claude/agents/name.md

`model`	Model (inherited by default)
`tools`	Restrict tools
`isolation: worktree`	Separate git worktree
`memory: project`	Persistent memory (user/project/local)
`permissionMode`	Dedicated permission mode
`description`	Criteria for Claude to invoke
`effort`	Sub-agent effort level

Invoke: @"name (agent)" | --agent name | SendMessage to: <id>

Builtin: Explore (Haiku), Plan, general-purpose, claude-code-guide, statusline-setup, code-reviewer.

6. MCP + Plugins

MCP: Integrations via .mcp.json. Tools: mcp__server__tool. OAuth RFC 9728 (Protected Resource Metadata) NEW.

# Add / List / Remove
claude mcp add playwright -- npx @playwright/mcp
claude mcp list
claude mcp remove playwright
# Import from Claude Desktop / JSON
claude mcp add-from-claude-desktop
claude mcp add-json '{"cmd":"..."}'
# MCP server exposed by Claude Code
claude mcp serve
# Debug
claude --mcp-debug

Transports: stdio | http | sse | ws

Size override: anthropic/maxResultSizeChars up to 500K/tool NEW

Plugins: Skills + agents + hooks + MCP + bin/ on PATH NEW. Auto-resolution of dependencies.

/plugin install name@marketplace

🖥 Computer Use — now in the CLI NEW

Claude sees and controls your desktop: screenshots, clicks, typing, scrolling, drag & drop. Since v2.1.86, available in the CLI (not just Desktop). macOS and Windows Desktop — Linux not supported. Multi-monitor. Pro/Max only. Enabled in Settings > General > Computer Use.

Security: request_access required. Allowlist per app. Screenshots filtered by compositing — only authorized apps are visible. systemKeyCombos separate grant.

Use: open a native app, click the UI, test changes, verify on iOS/macOS/GUI-only apps.

27 tools: screenshot, left_click, type, key, scroll, zoom, computer_batch (sequence in 1 call), open_application, mouse_move, double_click, right_click, left_click_drag, read/write_clipboard, wait, switch_display, etc.

👁 Monitor tool NEW

Built-in tool (v2.1.98) to spawn background watchers — tail logs, watch CI, auto-fix crashes. Each stdout line = notification in the conversation. Replaces Bash sleep loops. Pairs with /loop self-pacing.

🛡 Permission Modes

Mode	Behavior
Default	Asks permission for every tool
acceptEdits	Auto-approves edits, asks for Bash
Plan	Read-only. Explores & understands. Writes nothing.
Auto	AI classifies (fast + thinking path). Max + Opus 4.7 NEW
dontAsk	Skip prompts, ignore non-approved tools
bypass	Skip all checks (containers only)

Normal → Auto-Accept → Plan → Auto

Shift+Tab to cycle. --enable-auto-mode to unlock Auto. See Row 5 for classifier details.

Pro workflow: Plan Mode to explore → Normal to implement → Auto for long runs → Rewind if needed.

Pre-approve Tools

// .claude/settings.json
"allow": ["Read", "Bash(git *)", "Bash(npm run *)"]
"ask": ["Bash(npm publish *)"]
"deny": ["Bash(rm -rf *)"]

Order: deny → ask → allow (first match wins)

Syntax: Bash(npm run *), Read(/src/**), WebFetch(domain:x.com), mcp__server__*, Edit(*.ts)

April hardening: Bash escape of flags, /dev/tcp, env-var prefixes; PowerShell auto-approval matches Bash; sandbox.failIfUnavailable.

🔗 Hooks — Automation

Events (26)

Event	When
`SessionStart`	Session start / resume
`SessionEnd`	Session end
`UserPromptSubmit`	Before Claude processes the message
`PostUserPromptSubmit`	After submit — can override `sessionTitle` NEW
`PreToolUse`	Before a tool — can block (exit 2)
`PostToolUse`	After a successful tool — receives `duration_ms` NEW
`PostToolUseFailure`	After a failed tool
`PermissionRequest`	Permission dialog appears
`PermissionDenied`	After auto-mode denial (`retry: true`) NEW
`Stop`	Claude finishes its response
`StopFailure`	Turn ended by API error
`SubagentStart`	A sub-agent starts
`SubagentStop`	A sub-agent finishes
`TeammateIdle`	A teammate is about to become idle
`TaskCreated`	Task created via TaskCreate
`TaskCompleted`	Task completed
`PreCompact`	Before compaction — `exit 2` to block
`PostCompact`	After compaction
`Notification`	Claude sends a notification
`InstructionsLoaded`	CLAUDE.md or rules loaded
`ConfigChange`	A config file changes
`FileChanged`	A watched file changes
`CwdChanged`	Working directory changes (reload env)
`WorktreeCreate`	Worktree created
`WorktreeRemove`	Worktree removed
`Elicitation`	MCP server requests input
`ElicitationResult`	Response to an MCP elicitation

5 Handler Types

`command`	Shell command
`http`	POST to an endpoint
`prompt`	LLM single-turn (Haiku)
`agent`	Multi-turn sub-agent
`mcp_tool`	Direct MCP tool call NEW

// Conditional auto-format + duration_ms logging
"PostToolUse": [{
  "matcher": "Write|Edit",
  "if": "Edit(*.ts)",
  "hooks": [{
    "type": "command",
    "command": "npx prettier --write $CLAUDE_FILE_PATHS"
  }]
}]

Codes: 0 = success, 2 = block (stderr → Claude), other = error

if field: Filter by tool name AND arguments. E.g.: Bash(git *), Edit(*.ts), mcp__server__tool

Env vars: $CLAUDE_PROJECT_DIR, $CLAUDE_FILE_PATHS, $CLAUDE_TOOL_INPUT, $CLAUDE_SESSION_ID

hookSpecificOutput: {"sessionTitle": "..."} to name the session from a hook NEW

✨ Input Superpowers

Syntax	Action
`@file`	Reference files & folders in context
`@"agent"`	Mention a specific sub-agent
`!command`	Run a shell. E.g.: `! git status`
`Ctrl+V`	Paste screenshots, diagrams, images

Pipe Input

# Analyze logs
cat error.log | claude -p "find root cause"

# PR review
gh pr diff 42 | claude -p "security review"

# Commit message
git diff --staged | claude -p "conventional commit msg"

# Translate
cat README.md | claude -p "translate to Spanish"

Multi-directories & Worktrees

# Add directories
claude --add-dir ../lib --add-dir ../shared

# Isolated worktree
claude -w feature # isolated git branch

# Worktree + tmux
claude -w feature --tmux

Background & Remote

Ctrl+B to send to background. Claude continues while you work.

--remote "task" to launch an agent on claude.ai.

/remote-control to make the session remotely accessible.

Built-in Agent Tools

What Claude can call (used in allowedTools, if, hooks):

Filesystem	`Read`, `Write`, `Edit`, `NotebookEdit`, `Glob`, `Grep`
Shell	`Bash`, `BashOutput`, `KillBash` (+ `PowerShell` opt-in)
Web	`WebSearch`, `WebFetch`
Agents & Skills	`Agent` (Task), `Skill`, `ToolSearch`, `Monitor` NEW
Session / Meta	`TodoWrite`, `AskUserQuestion`, `ExitPlanMode`, `PushNotification`
MCP	`mcp__<server>__<tool>` (auto-prefixed)

Restrict: --tools Read,Edit,Bash(git *) · Disallow: --disallowedTools Bash

⚙ Configuration

Hierarchy (decreasing priority)

Enterprise	`/etc/claude-code/managed-settings.json`
CLI Args	Flags passed at launch
Local	`.claude/settings.local.json` (gitignored)
Project	`.claude/settings.json` (committed)
User	`~/.claude/settings.json` (global)

Config Commands

config listShow all settings config get keyView a value config set key valSet a value config add key valAdd to an array config remove key valRemove from an array

Environment Variables

`ANTHROPIC_API_KEY`	API key (overrides subscription)
`ANTHROPIC_MODEL`	Default model
`ANTHROPIC_SMALL_FAST_MODEL`	Model for lightweight tasks
`CLAUDE_CODE_EFFORT_LEVEL`	Effort level (low…xhigh/max/auto)
`CLAUDE_CODE_USE_BEDROCK`	Amazon Bedrock
`CLAUDE_CODE_USE_VERTEX`	Google Vertex AI
`CLAUDE_CODE_USE_MANTLE`	Bedrock via Mantle NEW
`CLAUDE_CODE_DISABLE_1M_CONTEXT`	Disable 1M context
`CLAUDE_CODE_DISABLE_AUTO_MEMORY`	Disable auto memory
`CLAUDE_CODE_NO_FLICKER`	1 = alt-screen (default v2.1.89), 0 = classic NEW
`CLAUDE_CODE_HIDE_CWD`	Hide cwd in startup logo NEW
`CLAUDE_CODE_CERT_STORE`	`bundled` to opt out of OS store NEW
`CLAUDE_CODE_SUBPROCESS_ENV_SCRUB`	Linux PID namespace isolation NEW
`CLAUDE_CODE_SCRIPT_CAPS`	Per-session script invocation limits NEW
`CLAUDE_CODE_PERFORCE_MODE`	Edit/Write fail read-only (`p4 edit` hint) NEW
`CLAUDE_CODE_USE_POWERSHELL_TOOL`	Enable/disable PowerShell tool
`ENABLE_PROMPT_CACHING_1H`	1-hour prompt cache TTL NEW
`FORCE_PROMPT_CACHING_5M`	5-min prompt cache TTL (test) NEW
`ENABLE_TOOL_SEARCH`	Enable ToolSearch on Vertex AI NEW
`OTEL_LOG_TOOL_DETAILS`	Emit full tool details in OpenTelemetry NEW
`DISABLE_TELEMETRY`	Opt-out telemetry
`DISABLE_AUTOUPDATER`	Disable auto updates
`DISABLE_UPDATES`	Block all update paths NEW
`CLAUDE_PROJECT_DIR`	Project root (injected in hooks)

Settings.json — Key Fields

`model`	Default model
`effort`	Default effort level
`permissions.{allow,ask,deny}`	Permission rules
`hooks`	Lifecycle hooks
`allowedMcpServers`	Whitelist MCP servers
`deniedMcpServers`	Blacklist MCP servers
`prUrlTemplate`	Custom code-review URL template NEW
`sandbox.failIfUnavailable`	Enforce sandbox availability NEW
`disableSkillShellExecution`	Block shell from skills/commands/plugins NEW
`showThinkingSummaries`	Thinking summaries (off in interactive v2.1.86) NEW
`statusLine.refreshInterval`	Re-run status line every N sec NEW
`theme`	Theme name or custom JSON

Advanced Permissions

// .claude/settings.json
{
  "permissions": {
    "allow": ["Read", "Bash(git *)"],
    "ask": ["Bash(npm publish *)"],
    "deny": ["Bash(rm -rf *)"]
  },
  "allowedMcpServers": ["playwright"],
  "deniedMcpServers": ["untrusted"]
}

📁 File Structure Map

Project Level (`.claude/`)

CLAUDE.md — Project memory, conventions settings.json — Shared settings (git) settings.local.json — Personal settings (gitignored) commands/ — Project slash commands skills/ — Skills (SKILL.md + references/) agents/ — Project sub-agents rules/ — File-scoped rules .mcp.json — Project MCP config (committable)

Global Level (`~/.claude/`)

CLAUDE.md — Applies to ALL projects settings.json — Global settings (persisted by /config) commands/ — Personal commands skills/ — Personal skills agents/ — Personal agents rules/ — Global rules themes/ — Custom JSON themes NEW plans/ — Plan files (/plan & /ultraplan) keybindings.json — Custom keybindings ~/.claude.json — User MCP config

Hierarchy (5 levels)

Policy > Organization > User global > User local > Project

⏪ Rewind & Checkpoints

Esc EscOpen rewind menu anytime /rewindSame function via command

Rewind Options

Option	Effect
Conversation	Go back in chat. Code stays.
Code	Restore files. Conversation stays.
Full Rewind	Restore conversation AND code.

Limits: Side effects (Bash, deleted files, installed packages) CANNOT be undone. Checkpoints only track file edits by Claude. Use Git.

Context Management

Startup budget ~6K tokens: system prompt, memory, env, MCP, CLAUDE.md, rules, skill descriptions.

Compaction: Auto at ~95% capacity. /compact [focus] triggers manually with a focus area. CLAUDE.md re-read from disk.

"Document & Clear" workflow: Save progress → /clear → Start fresh. Claude reads the summary.

Save tokens: Scoped rules, sub-agents for verbose tasks, --bare, context: fork on skills, opus[1m] (native on Opus 4.7).

Prompt caching: ENABLE_PROMPT_CACHING_1H (1h TTL) — /context shows cache hits (fixed % on Opus 4.7 in April). NEW

/resume: ~67% faster on large sessions (40MB+) since April. NEW

Write tool: +60% speedup in April 2026. MCP startup noticeably quicker. NEW

🤖 Auto Mode NEW

Permission classifier: the AI decides if each tool is safe or risky — no more dialogs for every Bash call.

Availability

Research preview	v2.1.83 (Week 13, March 23 2026)
GA on Max	v2.1.92+ (Week 15) with Opus 4.7
Team / Enterprise	Sonnet 4.6, Opus 4.6, Opus 4.7

Activation

--enable-auto-modeUnlock at launch Shift+TabCycle to the "Auto" pill /configPersist the setting

Classifier — 2 paths

Fast path: Haiku scans "safe / risky / uncertain". Thinking path: Opus/Sonnet takes over the uncertain cases to cut false positives.

Hook `PermissionDenied`

"PermissionDenied": [{
  "hooks": [{
    "type": "command",
    "command": "./scripts/request-allow.sh"
  }]
}]
// stdout {"retry": true} => retry the tool

Default rules: claude auto-mode defaults. Effective config: claude auto-mode config.

No conflict: Plan Mode stays prioritized when both are active (April fix).

🧠 Models & Effort Levels

Available Models

Model	ID	Notes
Opus 4.7	`claude-opus-4-7`	1M context, 64K-128K output, adaptive thinking NEW
Opus 4.6	`claude-opus-4-6`	Fast mode with 1M context (April)
Sonnet 4.6	`claude-sonnet-4-6`	Fast & balanced, default
Haiku 4.5	`claude-haiku-4-5-20251001`	Ultra-fast, exploration, classifier

Adaptive Thinking NEW

On Opus 4.7, extended thinking is decided by the model — no more manual budget_tokens. Inline indicators: "still thinking", "thinking more", "almost done thinking".

Effort Levels

low medium high xhigh max auto

`low`	Simple questions, exploration
`medium`	Common usage
`high`	Default for API-key / Bedrock / Vertex / Team / Enterprise
`xhigh`	100K tokens — default for Opus 4.7, MRCR v2 = 71% NEW
`max`	Maximum budget — hard problems
`auto`	Research preview: classifier decides

Invoke

--effort xhighAt launch /effortInteractive slider (↑↓) NEW "ultrathink"One-shot max effort in the prompt Option+TToggle thinking

🌍 Platforms

CLI

macOS	Native installer (npm or `curl`)
Linux	Native installer — PID namespace isolation NEW
Windows	Native installer (WSL no longer required)

Desktop App NEW

Redesign April 14, 2026:

Multi-session sidebar
Drag-&-drop pane layout
Integrated terminal
File editor + HTML/PDF preview
Rebuilt diff viewer
Computer Use macOS & Windows

/desktop from the CLI to transfer a session to the app.

Web — claude.ai/code NEW

Research preview for Pro/Max. Cloud infrastructure, persistent sessions on mobile. /teleport to resume a web session locally.

IDE

VS Code	Multiple agentic extensions at once (Copilot + Claude Code + Codex) NEW
JetBrains	Beta plugin — improved diff viewer
/ide	Manage active integrations

GitHub & GitLab

/install-github-app — GitHub Actions & Code Review. --from-pr accepts GitHub, GitLab, Bitbucket, GitHub Enterprise. GitHub MCP server (mcp__github__*).

Third-party: "3rd-party platform" option on the login screen. Wizards claude bedrock setup / claude vertex setup.

🎯 Pro Workflow — Getting the Most Out of Claude Code

Starting a Project

1.cd project && claude 2./init to generate CLAUDE.md 3.Edit CLAUDE.md → conventions & commands 4./powerup to discover features NEW 5.Start coding!

Plan → Execute Loop

Shift+Tab × 2 → Plan Mode → Describe the task

Validate → Shift+Tab → Normal or Auto → "implement the plan"

/ultraplan to offload planning to the cloud, review in the browser, execute locally. NEW

Manage Context

/compact [focus]Compact with a focus area /clearUnrelated tasks? Reset opus[1m]Large codebases (native on Opus 4.7) --bareQuick questions, skip startup /focusCondensed view (quick read) NEW

Watchers & Automation

MonitorBackground watcher, tail logs NEW /loopSelf-pacing (omit interval) NEW Hook PostToolUseAuto-format after Edit

Prompting Techniques

Precisecat error.log | claude -p "root cause" Verifiable"After the fix, run npm test" IncrementalBreak into steps, validate each one Correct earlyCtrl+C if wrong direction AdaptiveOn Opus 4.7, let the model decide thinking NEW

Advanced Patterns

Worktrees	`claude -w feat --tmux`
Multi-dirs	`--add-dir ../lib`
Background	`Ctrl+B`
Agent Teams	`--teammate-mode auto`
Remote	`--remote "task"`
Loop	`/loop 5m /test` or self-pacing
Schedule	`/schedule` for cloud cron
Auto Mode	Max+Opus 4.7: classifier handles permissions NEW
From PR	`--from-pr 42` (GitHub/GitLab/Bitbucket)

Headless / CI

claude -p --max-budget-usd 2 \
  --max-turns 5 \
  --effort xhigh "fix failing test"

OpenTelemetry: tool_use_id, tool_input_size_bytes, command_name, effort attrs NEW

Learning loop: Claude makes a mistake → you correct it → promote to a MUST rule → Claude never repeats it. This is compound learning.

📝 Customize Claude Code — Create Your Extensions

📋 Slash Command — `.claude/commands/name.md`

---
description: Review staged changes
allowed-tools: Read, Bash(git *)
model: sonnet
disable-model-invocation: true
argument-hint: "[focus area]"
---
Diff: !`git diff --staged`
Review: $ARGUMENTS | $1, $2 positional

You invoke it: /name. Personal: ~/.claude/commands/

🧠 Skill — `.claude/skills/name/SKILL.md`

---
name: deploy
description: "Deploy to production (1536 chars max)"  # NEW
context: fork          # isolated sub-agent
paths: "src/**/*.ts"    # activate by glob
effort: high           # force extended thinking
user-invocable: true  # false = Claude only
allowed-tools: Read, Bash(git *)
model: sonnet
---
Instructions... + [ref](references/guide.md)

Claude decides when to invoke it (auto-discovery via description). /skills sorts them by token cost.

🤖 Sub-agent — `.claude/agents/name.md`

---
name: security-reviewer
description: "Audit code for vulnerabilities"
model: sonnet
tools: Read, Grep, Glob
memory: project        # user / project / local
isolation: worktree    # separate git worktree
permissionMode: plan
---
You are a security specialist...

Invoke: @"name (agent)" | --agent name | natural language

📐 Rule — `.claude/rules/name.md`

---
paths: ["src/**/*.tsx", "src/**/*.ts"]
---
MUST use TanStack Query for data fetching.
SHOULD prefer named exports over default.
MAY use inline styles for one-off components.

Auto-loaded when Claude touches a file matching the glob.

🔗 Hook — `.claude/settings.json`

"hooks": {
  "PostToolUse": [{
    "matcher": "Write|Edit",
    "if": "Edit(*.ts)",
    "hooks": [{
      "type": "command",
      "command": "npx prettier --write $FILE"
    }]
  }]
}

Deterministic — always runs (unlike CLAUDE.md ~80%).

🔌 MCP Server — `.mcp.json`

# Add via CLI
claude mcp add playwright -- npx @playwright/mcp
# Or in .mcp.json
{ "mcpServers": { "pg": {
  "type": "stdio",
  "command": "npx",
  "args": ["@mcp/server-postgres"]
}}}

Essentials

Navigation & Display

Editing & Model

Vim mode NEW

Quick Prefixes

Confirmation Dialogs

Session & Context

Model & Configuration

Project & Analysis

Extensions & Plugins

Advanced & Remote

System & Account

Session Startup

Model & Reasoning

Context & Directories

Permissions

Output & Formats

System Prompt

Advanced & Remote

CLI Subcommands

1. CLAUDE.md — System Prompt

2. Rules — Scoped Rules

3. Custom Commands — Shortcuts

4. Skills — Complex Workflows

5. Sub-agents — Parallel Specialists

6. MCP + Plugins

🖥 Computer Use — now in the CLI NEW

👁 Monitor tool NEW

Pre-approve Tools

Events (26)

5 Handler Types

Pipe Input

Multi-directories & Worktrees

Background & Remote

Built-in Agent Tools

Hierarchy (decreasing priority)

Config Commands

Environment Variables

Settings.json — Key Fields

Advanced Permissions

Project Level (.claude/)

Global Level (~/.claude/)

Hierarchy (5 levels)

Rewind Options

Context Management

Availability

Activation

Classifier — 2 paths

Hook PermissionDenied

Available Models

Adaptive Thinking NEW

Effort Levels

Invoke

CLI

Desktop App NEW

Web — claude.ai/code NEW

IDE

GitHub & GitLab

Starting a Project

Plan → Execute Loop

Manage Context

Watchers & Automation

Prompting Techniques

Advanced Patterns

Headless / CI

📋 Slash Command — .claude/commands/name.md

🧠 Skill — .claude/skills/name/SKILL.md

🤖 Sub-agent — .claude/agents/name.md

📐 Rule — .claude/rules/name.md

🔗 Hook — .claude/settings.json

🔌 MCP Server — .mcp.json

Basic Commands

Models & Effort

Power Moves

Project Level (`.claude/`)

Global Level (`~/.claude/`)

Hook `PermissionDenied`

📋 Slash Command — `.claude/commands/name.md`

🧠 Skill — `.claude/skills/name/SKILL.md`

🤖 Sub-agent — `.claude/agents/name.md`

📐 Rule — `.claude/rules/name.md`

🔗 Hook — `.claude/settings.json`

🔌 MCP Server — `.mcp.json`