25 Claude Code Tips, Shortcuts, and Power Features

Claude Code rewards depth: the default chat loop works on day one, but the operators who get 10x from it are using a different toolset — prefixes, modes, session surgery, and a few flags. Here are the 25 tips that pay off most, grouped by what they speed up.

Input: the four prefixes

1. ! runs shell directly. !npm test executes immediately and puts the output in context — no asking, no narration. The fastest way to feed Claude the ground truth before your next ask.

2. @ references files. Refactor @src/auth/session.ts to use the new client attaches the file. @ a directory for its listing; mention several files in one message to scope a task precisely.

3. # saves a memory. # this repo uses pnpm, never npm persists the fact to memory instead of dying with the session.

4. / is the command palette. Everything below with a slash is discoverable by typing / — and your own custom commands join the list.

Flow control

5. Esc interrupts. Claude stops mid-action; context survives. Watching it head somewhere wrong for 30 seconds and waiting politely is the most common wasted minute in agentic coding.

6. Double-Esc rewinds the conversation. Jump back to a previous message, edit it, rerun from there — fix the prompt instead of patching its consequences.

7. /rewind undoes a wrong turn. Roll code and conversation back to a checkpoint when an approach failed. Cheaper than asking the agent to un-dig its hole.

8. Queue your next message. You can keep typing while Claude works — queued messages deliver when the current step finishes. Batch your guidance instead of waiting at the spinner.

9. Shift+Tab cycles permission modes. default → acceptEdits → plan without touching settings. Trust rising? acceptEdits. Stakes rising? plan.

10. Plan mode before big changes. In plan mode Claude explores read-only and proposes; nothing is edited until you approve. The cheapest insurance in the product.

Sessions

11. claude --continue reattaches. The most recent session in this directory, instantly. /resume opens the picker when you need an older one.

12. Sessions are per-directory. Each project — and each git worktree — keeps its own history. That's the mechanism behind clean parallel sessions.

13. /clear between unrelated tasks. Stale context degrades output quality and costs tokens. Clear freely — the old session stays reachable via /resume.

14. /compact with instructions. Don't wait for auto-compaction: at a milestone, /compact keep the failing tests and the migration plan compresses on your terms.

15. /context shows the bill. A visual map of what's consuming the window — usually the moment you discover three MCP servers you stopped using.

16. /export and /copy. Export the transcript to a file, or copy the last response — for the PR description, the ticket, the teammate who asked "what did it do?"

Models and thinking

17. Switch models per task. /model mid-session: Opus-tier for the gnarly design, Sonnet for the long middle, Haiku for the mechanical sweep. See Choosing the Right Model.

18. opusplan splits the difference. Plan with Opus, execute with Sonnet — strong thinking where it matters, efficiency where it doesn't.

19. "ultrathink" for the hard one. Include it in a prompt to max out reasoning on that turn; toggle extended thinking session-wide with Option+T / Alt+T, and view the reasoning with Ctrl+O.

20. Paste images. Screenshots of broken UI, error dialogs, whiteboard diagrams — paste straight into the prompt (Ctrl+V; some terminals vary). A screenshot of the bug beats three sentences about it.

Automation: make it yours

21. Custom slash commands. A Markdown file in .claude/commands/ becomes /your-command; $ARGUMENTS interpolates what you pass. Any prompt you've typed twice deserves the third time to be one word — steal ideas from the commands directory.

22. Pipe into headless mode. cat error.log | claude -p "find the root cause" — Claude Code as a Unix filter. -p is also the CI entry point, with JSON output when scripts consume the answer.

23. Hooks make rules deterministic. Format-after-edit, block-protected-paths, notify-when-waiting — hooks run every time, no model memory required.

24. /statusline and /output-style. Put model, branch, and context usage in your status bar; tune output verbosity to taste. Small, but you look at it all day.

25. /doctor first, then debug. Installation weirdness, version drift, broken MCP connections — /doctor diagnoses the boring causes before you burn time on interesting theories. Its sibling /usage shows spend and plan limits before the invoice does.