Claude Code Plan Mode: see the plan before it touches your code

Plan mode is a read-only mode in Claude Code. You toggle it on, and Claude explores your codebase and writes up a full plan for the change — without editing a single file or running anything that changes state. You read the plan, steer it if the approach is off, and only then approve it. That's when Claude actually starts building.

TL;DR: press Shift+Tab to cycle into plan mode. Claude investigates and proposes a step-by-step plan but changes nothing. Review it, tweak your ask if needed, then approve — Claude exits plan mode and executes. Best for big, risky, or multi-file work where you want to approve the approach before any code moves.

---

Why it exists

The default failure mode of an eager agent: you give a one-line ask, and it runs off and edits fifteen files before you realize it misread you. Now you're reverting a pile of changes by hand and re-explaining what you actually meant.

Plan mode flips the order. Claude does all the reading first — maps the files, spots the edge cases, and hands you the approach as a plan. You catch the misunderstanding while it's still just words on the screen, not a diff across your repo.

---

How to turn it on

Press Shift+Tab. In the Claude Code prompt, Shift+Tab cycles through the input modes. Keep tapping it until the indicator reads "plan mode":

normal  →  auto-accept edits  →  plan mode

You'll see the mode shown near the prompt. Now type your request as usual — Claude will research and respond with a plan instead of jumping straight to edits.

When you're happy with the plan, approve it. Approving exits plan mode and Claude carries out the plan for real. If the plan is wrong, you don't approve — you adjust the ask and let it re-plan.

---

What it will and won't do

The whole point: it investigates and drafts, but changes nothing until you say go.

---

A concrete workflow

"Add email + password auth to the app. Use plan mode first."

1. You press Shift+Tab until it shows plan mode, then send the ask.
2. Claude reads routes/, db/schema.ts, and the existing UI, then returns a plan: add the /auth route, extend the schema, wire a login form, add session handling — listed step by step, with the files it'll touch.
3. You spot that it planned a new sessions table when you already have one. You reply: "we already have sessions — reuse it." Claude re-plans.
4. The approach now looks right. You approve. Claude exits plan mode and executes the plan, editing the files it laid out.

You steered the approach in two sentences — before a single line of code changed.

---

When to reach for it

• Big refactors — renames and moves that ripple across dozens of files.
• Risky changes — auth, payments, migrations, deletes; anything expensive to undo.
• Unfamiliar codebases — when you don't yet know the blast radius of a change.
• When the how matters — you want to approve the approach, not just the outcome.

For a quick, single-file edit you already understand, skip it — plan mode is overhead you don't need. It earns its keep the moment a task is wide or hard to reverse.

---

Tips

• Read the plan like a code review. It's your cheapest chance to catch a wrong approach. Push back in plain language — "do X instead of Y", "don't touch Z".
• Front-load constraints in the ask. "Keep the existing API", "no new dependencies", "match the current folder layout" — the plan will respect them.
• Re-plan freely. If the approach is off, adjusting your ask and re-planning costs nothing. Cheap iterations here beat expensive reverts later.
• Pair it with a broad ask. Plan mode shines on "here's the goal" prompts where the path isn't obvious. It turns a vague request into a reviewable plan you can trust.

---

Guide by StackScout · more free AI-dev guides, skills & repo picks → stackscout.app