Claude Code cheat sheet.

Install the CLI globally. Requires Node ≥ 20.

Launch the interactive TUI in the current directory.

Sign in with your Anthropic account (browser flow).

Print CLI version. Update often — releases ship weekly.

Self-update the CLI to the latest release.

Diagnose install / permissions / MCP issues.

Send the current prompt.

Newline inside the prompt (multiline).

Cancel the running turn (interrupt tool calls).

Exit the session cleanly.

Cancel autocomplete / close popups.

Cycle through prior prompts in this session.

Prefix a line to run it as a raw shell command (result lands in context).

Prefix to send a system-style note to the model without asking for action.

Full command list + docs pointers.

Reset the current context window (keeps CLAUDE.md loaded).

Manually compact history when you're low on tokens.

Open the settings UI (theme, model, permissions).

Switch model mid-session (Opus / Sonnet / Haiku).

Toggle Fast mode on Opus 4.6/4.7 — faster output, same model.

Create a CLAUDE.md scaffold from the current codebase.

Review a pull request.

List + manage MCP servers for this session.

Inspect + edit the hook settings for this project.

List available subagents and their tool policies.

Set a session-scoped goal with a Stop hook that keeps you focused.

Run a prompt or slash command on a recurring interval.

Schedule a remote agent (routine) on a cron.

Push the current branch to origin with a pre-push summary.

User-scope preferences — applies to every project.

Project-scope — checked into the repo, applies to all contributors.

Local-only overrides — .gitignore'd, per-machine.

Under ~200 lines. Model reads it every turn — bloat costs tokens forever.

Bulleted 'always / never' rules land better than paragraphs of context.

The skill definition — name, description, args, body.

The router. Write it as trigger phrases, not a summary.

Meta-skill to create, edit, test, and optimize other skills.

Fully-qualified name for plugin-namespaced skills.

Everything after the skill name is passed as one args string.

Fires before a tool runs — can block or mutate arguments.

Fires after — good for lint, format, dry-run gates.

Injects a system reminder each time the user submits.

Blocks the session from ending until a condition holds (used by /goal).

Set up / tear down env, log, or announce.

Project-scope hook config (checked into the repo).

Local-only overrides — not committed.

Skill that safely edits settings.json for you.

Register an MCP server (stdio or HTTP).

List configured MCP servers and their status.

Deregister a server.

Naming convention for MCP-provided tools in the session.

Enumerate readable resources exposed by a server.

Fetch a resource by URI.

Subagent definition — tools, model, description.

Fast read-only search subagent (files + grep + read).

Architect subagent — designs step-by-step plans.

Wildcard — research or multi-step tasks.

Launch a subagent from the main loop; get back a summary.

Fire-and-forget parallel subagent (notified when done).

Run in an isolated git worktree; auto-cleanup on no-change.

Switch into plan mode (usually via UI or a slash command).

Commit to a plan and get approval before implementing.

Ask for a plan before letting Claude touch files.

Send Plan-mode reads to subagents, keep main context clean.

Run a slash command every interval, e.g. `/loop 5m /watch-ci`.

Self-paced loop — Claude decides when to fire next.

Create a remote scheduled agent (routine) on a cron.

Tool-level control over scheduled agents.

In-loop tool to pick the next wake time in dynamic loops.

Never --amend or force-push unless the user explicitly asks.

Hook failures are fixed at the root, not bypassed.

Prefer `git add file1 file2` over `git add .` to avoid secrets.

Use gh for PR/issues/checks — never guess GitHub URLs.

Commit trailer the CLI appends by default.

Open the permissions UI.

Auto-accept file edits (careful — pair with plan mode).

Block all edits (read-only mode).

Explicit allowlist of shell / tool patterns in settings.json.

Skill that scans your transcript and adds safe allowlist entries.

The SDK Claude Code itself is built on.

Anthropic-hosted agent runtime (April 2026 launch).

Skills, hooks, MCP all portable between local + managed.

Master agent skill for learning Managed Agents end-to-end.

Paid model first, then :free — free models 429 under load.

Add --non-interactive to CLI shells; --yes alone hangs.

Set S3_REGION explicitly or S3 mis-targets us-east-1.

Don't run Playwright headed via WSLg — crashes low-RAM laptops.

Central cluster near 500/500 — fold to FLEET shape, don't add new collections.