Starts an interactive session in the current folder.

Go into your project (cd my-project) then type claude.

Starts by asking a question straight away.

claude "explain this project structure in three sentences".

Resumes the most recent conversation in the current folder.

Closed the terminal yesterday? claude -c picks up where you left off.

Lists your past sessions so you pick one to resume.

Short alias: claude -r. Handy to find a specific task.

Updates Claude Code to the latest version.

Run it now and then; the tool evolves fast.

Diagnoses the installation and configuration.

Odd behaviour? Run claude doctor first.

Prints the installed version number.

Useful to confirm an update or report a bug.

Shows the list of all available commands.

The built-in cheat sheet: no need to memorise commands.

Starts from an empty context, for an unrelated task.

Switching from CSS to an import script? /clear tidies up.

Summarises the conversation to free context, keeping the essentials.

Long session getting fuzzy? /compact keeps what matters.

Visualises how full the session context is.

To tell whether it is time to /compact or /clear.

Switches model: fast and cheap, or more powerful.

Small task → a fast model; tricky problem → a powerful one.

Shows your usage and costs (alias: /cost).

To keep an eye on the bill when paying per API use.

Opens settings (theme, editor mode, preferences).

Turn on vim editor mode for input, for instance.

Returns to an earlier conversation without leaving the session.

Like claude --resume, but from inside the current session.

Rolls back code AND conversation to a checkpoint.

A change you dislike? /rewind cleanly reverts it.

Exports the conversation as plain text.

Handy to archive or share a session.

Cycles modes: normal → auto-accept edits → plan.

The key shortcut: switch to plan mode in one keystroke.

Starts straight in plan mode: read-only, then a plan to approve.

Ideal before any refactor: Claude proposes without touching anything.

Manages the allow / ask / deny rules for tools.

Allow npm test once and for all; deny reading .env.

Views and configures hooks (commands run automatically).

E.g. run prettier after every file Claude edits.

Skips all permission prompts. Reserve for isolated environments.

Fine in a throwaway CI container; never on your machine.

Generates a starter CLAUDE.md by analysing the project.

First reflex on a new project: run /init, then refine.

Opens and edits the CLAUDE.md memory files.

To add a team convention or a test command.

Creates and manages specialised subagents (separate context).

E.g. a read-only "code reviewer" subagent.

A markdown file here (or a skill) creates a custom /name command.

Create .claude/commands/review.md to get /review.

Adds a local MCP server (a capability plugged into Claude).

claude mcp add playwright -- npx @playwright/mcp@latest (browser).

Adds a remote MCP server, reachable via a URL.

For a hosted service (Sentry, Notion…) rather than a local one.

Lists your MCP servers and their status.

claude mcp get <name> for details, claude mcp remove <name> to remove.

Shows MCP server status and starts authentication (OAuth).

A server needs sign-in? Type /mcp and follow the browser.

Wires your repo: mentioning @claude in an issue/PR triggers the agent.

"@claude fix this bug" in a comment, and the GitHub action runs.

At the start of a line: runs a command or a skill.

Type / to reveal the full, filterable list.

Runs a shell command directly, without going through Claude.

! git status shows the status and adds it to the conversation.

References a project file, with autocomplete.

Explain @src/auth.js: Claude reads the file itself.

Saves the following instruction into the CLAUDE.md file.

# Always run npm run lint before a commit → saved for good.

Interrupts Claude mid-action, keeping the work done so far.

Going the wrong way? Esc, then rephrase.

Clears the current draft, or opens the rewind menu.

On empty input, double Esc steps back through history.

Recalls your previous messages to reuse or edit them.

Reuse a request without retyping it all.

Completes a filename, or accepts the greyed-out suggestion.

After @ or /, Tab speeds up typing.

Pastes an image into the prompt (Claude reads images).

A screenshot beats ten lines of description. Alt+V if Windows grabs Ctrl+V.

Clears the current input; twice in a row, quits Claude Code.

Started a long request and changed your mind? Ctrl+C.

Toggles the detailed transcript view: every tool call.

To see exactly what Claude did, step by step.

Moves a long command to the background so you keep working.

A dev server running? Background it and move on.

"Print" mode: one answer, then exits. The heart of headless mode.

cat errors.log | claude -p "explain the root cause".

Returns structured output a script can parse (with jq, say).

claude -p "summarise" --output-format json | jq -r ".result".

Strict allowlist of tools permitted without confirmation.

Essential in CI: Claude can ONLY do what you list.

Limits the agent’s number of turns (a circuit breaker).

E.g. --max-turns 10 to cap the cost of an automated task.

Generates a long-lived token for CI and scripts.

Store it in your pipeline secrets, never in plain text.