Plan Mode in Claude Code: it doesn't make Claude smarter — it forces you to think

TL;DR Shift+Tab cycles to plan, or /plan <prompt> runs a single turn in plan. Claude explores, proposes a step-by-step plan, doesn't touch a single file. I combine it with /evaluate before implementing — and sometimes after too, to check what Claude did against the original plan. Plan mode doesn't give Claude superpowers; it forces you to actually understand what's about to change.

Plan Mode is one of the first steps toward working with an AI agent with any structure at all. The myth is that activating Plan Mode flips on superpowers — that Claude becomes smarter, more careful, more "himself." The reality is more mundane and more useful: what Plan Mode does is force Claude to break the task into the smallest possible atomic actions. And by forcing it to do that, it forces you to understand it. That's why it feels magical when it isn't. It's just the power of planning before acting — the same one humans get from it.

Technically, Plan Mode is a read-only permission mode: Claude reads files, runs read-only shell commands (grep, ls, find, cat), and writes a plan. It can't touch a single file until you explicitly approve the plan.

Result:

> /plan add rate limiting to the auth endpoints

⠋ Exploring codebase...
⠋ Reading src/auth/*.ts
⠋ Reading src/middleware/*.ts

Here's my plan for adding rate limiting:

1. Install express-rate-limit
2. Create src/middleware/rateLimit.ts with two tiers:
   - Strict: /login, /register, /reset-password (5 req/15min)
   - Standard: /me, /refresh (100 req/15min)
3. Wire into src/app.ts before auth routes
4. Add RATE_LIMIT_WINDOW and RATE_LIMIT_MAX to .env.example
5. Update README.md with the new env vars

Ready to proceed?
  ❯ Approve and start in auto mode
    Approve and accept edits
    Approve and review each edit manually
    Keep planning with feedback
    Refine with Ultraplan

Four ways to enter Plan Mode

1. Shift+Tab — cycle to plan mid-session

Press Shift+Tab to cycle through modes: default → acceptEdits → plan. The status bar shows the active mode. For the other three modes (auto, dontAsk, bypassPermissions), see the 6 permission modes.

2. /plan <prompt> — single turn only

Prefix a single prompt with /plan to enter plan just for that turn:

> /plan refactor the payment retry logic to use exponential backoff

Useful when most of the session is in acceptEdits but this particular change needs planning.

3. --permission-mode plan — start the session in plan

claude --permission-mode plan

The whole session starts in plan. Useful when you're exploring an unfamiliar repo or have to make an architecture call before touching anything.

4. defaultMode — plan by default in this project

// .claude/settings.json
{
  "permissions": {
    "defaultMode": "plan"
  }
}

Every session you start in this project begins in plan. Recommended for critical repos: production, payment systems, code that touches user data.

The workflow that actually works

When the plan is ready, Claude offers five options. The first four approve; the fifth refines:

Before approving, press Ctrl+G to open the plan in your default editor. Annotate steps, delete the ones you don't like, reorder operations, save. Claude uses the edited plan.

My personal flow: instead of approving directly, I run the plan through /evaluate first. The evaluator checks each step against the real codebase — if Claude says "I'll modify src/middleware/auth.ts line 42", /evaluate verifies that line exists and that the change makes sense. Sometimes I also run /evaluate after implementing, against the original plan, to catch silent drift. Plan → /evaluate → Implement → /evaluate. Sounds paranoid until it prevents the first disaster.

Even to change a button color?

Yes, even then. An unfiltered question can end up touching files you didn't expect — a shared util, a global theme, a test snapshot. Plan mode shows you what Claude plans to touch before it does. Five seconds of "ah, it's going to modify 8 files, not 1" save you 20 minutes of reverting.

The cost of planning is low. The cost of not planning can be anything.

Reference

Official docs: Choose a permission mode