Claude Code Cheat Sheet

Start an interactive Claude Code session in the current directory.

Continue the most recent conversation in the current directory.

Resume a conversation by session ID or name, optionally with a new prompt.

Show the available built-in commands and usage hints. Type / to filter the command menu.

Switch the AI model. With no argument, open the picker; supported models can also adjust effort with the arrow keys.

Start a new conversation with empty context. An optional name keeps the old session easy to find in /resume.

Summarize the current conversation to free context while keeping the same session. Add focus instructions when needed.

Initialize a project guide by creating a starter CLAUDE.md for the repository.

Resume a conversation by ID or name, or open the session picker. Alias: /continue.

Open Settings on the Status tab to inspect version, model, account, connectivity, and setting sources.

Exit the CLI. /quit is an alias of /exit; Ctrl+D also sends the exit signal.

Run a non-interactive prompt for scripts or CI. Pair it with an output format when a machine-readable result is needed.

Persistent project context: architecture, conventions, test commands, constraints, and instructions Claude should follow every session.

User settings apply everywhere; project settings are shared; local settings are gitignored overrides. Managed policy sits above them.

Set the initial model in settings, use --model for a launch override, or use /model during a session.

Allow, ask, or deny tool calls with specific rules. Deny rules take precedence; review active sources with /status and edit via /permissions.

Change how Claude responds. Built-in styles include Default, Explanatory, Proactive, and Learning; custom styles live in output-styles/.

Split reusable project rules into focused Markdown files. Keep personal project notes in CLAUDE.local.md; use @path/to/file to import context.

MCP is an open protocol for exposing tools, resources, and prompts to an AI client. In Claude Code, each server adds a reviewed connection to an outside system.

Add a remote HTTP server. HTTP is the recommended transport for cloud services and supports OAuth when the provider offers it.

Run a local MCP process. The double dash separates Claude Code options from the server command and its arguments.

List configured servers, inspect one server, or remove it. Inside a Claude Code session, use /mcp for status and authentication.

Local is the default and stays tied to you and one project. Project scope writes shared config to .mcp.json; user scope is available across your projects.

These provider endpoints are documented examples. Always review the provider’s current setup and permissions before connecting.

Pause and retry with backoff. Check account usage, plan or API billing, provider status, and whether multiple sessions are sharing the same limit.

Run /context to find the biggest consumers, then /compact with a focus. Start a new task with /clear and summarize only what matters.

Inspect the exact tool rule in /permissions. Check the working directory and the setting scope; a managed deny cannot be overridden by a project file.

Check the endpoint, transport, headers, proxy, and local network. For MCP, inspect health with claude mcp get and re-authenticate if needed.

Check whether Claude Code is authenticated, then sign in again. If you use Console billing, select the console login flow; do not commit credentials.

Command availability can vary by Claude Code version, plan, platform, and environment. Check your version and the current command menu before automating.