Cheat Sheet: Keyboard Shortcuts & CLI Flags

Claude Code is driven from two places, and this whole cheat sheet splits along that line.

• Keyboard shortcuts live inside a running session. They change how you steer Claude mid-flight: switch permission modes, interrupt, rewind to a checkpoint, search history, peek at the transcript. You press them at the prompt.
• CLI flags live outside a session — on the claude command you type in your shell to start one. They decide what kind of session you get: which model, which permission mode, headless or interactive, scripted with a budget, or resumed from before.

So there are really two questions this page answers fast: "I'm in a session — what key does the thing?" and "I'm launching — what flag sets it up?" Skim to the right table and copy the exact form.

One setup detail gates a chunk of the shortcuts on macOS: many of them use Alt/Option as a Meta key, and macOS terminals don't treat Option that way out of the box. If your Option+P or Option+T does nothing, you haven't broken anything — you just need the Option-as-Meta setup below. Do that once and the whole Alt/Option row lights up.

The content here is the verbatim reference from the official interactive-mode and CLI docs. Bookmark it; you'll come back.

These work at the prompt inside a running session. The ones in the Alt/Option family need the macOS Option-as-Meta setup first.

The two you'll wear out first are Esc and Esc Esc. Esc is your steering wheel — interrupt the moment Claude heads the wrong way, type a correction, and it picks up from there without a restart. Esc Esc is your undo — every user prompt and every file edit is checkpointed, so you can rewind code, conversation, or both.

Inserting a newline (a multi-line prompt without submitting) depends on your terminal. Pick the row that works for you:

If Shift+Enter doesn't make a newline in your terminal, run /terminal-setup inside Claude Code — it configures the terminal so the native binding works.

Two prompt prefixes change how a line is interpreted, and they only fire at the start of the line:

Use ! when you want a fact in front of Claude without asking it to do anything — !git status, !ls -la, !cat package.json. The output lands in context for the next thing you ask. Use # to jot a durable note — # we deploy from the release branch, never main — which gets written to your project's auto-memory file.

On macOS, the Option key normally types special characters (Option+P gives π, not a Meta keystroke). Claude Code's Alt/Option shortcuts — Option+P (switch model), Option+T (thinking), Option+Enter (newline), and the readline keys Alt+B/Alt+F/Alt+Y/Alt+M — need Option to act as a Meta key. Until you flip this setting, those shortcuts silently do nothing. This is the single most common "why doesn't this shortcut work?" on macOS.

Configure it once per terminal:

That VS Code key, in your settings.json:

{
  "terminal.integrated.macOptionIsMeta": true
}

After setting it, restart the terminal (or open a new tab) and try Option+P — the model switcher should appear. If it still types a special character, you edited the wrong profile or the wrong settings scope.

These shape an everyday claude invocation — the session you start by hand.

The trio you'll lean on daily: -p for one-shot headless answers (great in pipes — tail -200 app.log | claude -p "find anomalies"), -c to jump straight back into your last conversation here, and -r to pick a specific older one. Add --fork-session when you want to branch off an old session without disturbing it.

These pair with -p to make Claude safe and predictable inside scripts, CI, and automation — bounded spend, structured output, and a controlled prompt.

The golden rule for any unattended run is scope it three ways: cap spend with --max-budget-usd, cap iterations with --max-turns, and whitelist exactly what may run unprompted with --allowedTools. For structured pipelines, --json-schema plus --output-format json gives you output you can pipe into jq.

And for system-prompt edits, always prefer --append-system-prompt over replacing the whole prompt — appending keeps Claude's built-in tool knowledge and safety guidance, where a full replacement (--system-prompt) throws them away.