Claude Code Plan Mode: Approve the Plan Before It Builds

Plan mode is a Claude Code setting where Claude researches your project and writes out what it intends to do, then waits — no files change until you approve the plan. Press Shift+Tab in the terminal, or pick Plan from the mode selector in the desktop app.

If you can’t read code, every edit an AI makes is a leap of faith. A plan is written in English. It’s the one thing in this whole process you’re fully qualified to review — and reviewing it takes two minutes.

What plan mode actually does

Anthropic’s docs put it plainly: “Claude reads files, runs shell commands to explore, and writes a plan, but does not edit your source.” Edits stay blocked until you approve.

So in plan mode Claude will:

• Read your files and search the project to understand what’s already there.
• Write out the changes it intends to make, in what order, and why.
• Stop and wait for your answer.

While it’s on, the status bar shows ⏸ plan mode on. Press Shift+Tab again to leave without approving anything.

Two honest caveats. Claude can still run commands to explore your project while it plans: when auto mode is available on your account, a second model reviews those commands for you, and otherwise anything outside the built-in read-only set stops and asks you first. And the block on editing has one documented exception — sessions started with bypass permissions available, a setting you have to turn on deliberately.

Why this matters more if you don’t code

A developer reviews the diff — the line-by-line list of what changed. You can’t, and pretending otherwise is how founders end up with an app they don’t understand.

You can review a sentence like “add a subscriptions table and charge the user’s card on signup.” If that’s wrong, you catch it while it costs one sentence to fix, instead of after twelve files and a broken checkout.

Reading the plan also closes the gap between what you said and what Claude heard. Most bad AI builds start in that gap, with a reasonable app that happens to be the wrong one. Spot the mismatch in the plan and you spot it while it’s still one paragraph.

How to turn plan mode on

Pick whichever surface you already use:

1. Desktop app — in the Code tab, use the mode selector next to the send button and choose Plan. This is the no-terminal path; if you haven’t set up Claude Code yet, start with our non-coder’s guide.
2. Terminal — press Shift+Tab to cycle permission modes until the status bar reads ⏸ plan mode on. To start there, launch with claude --permission-mode plan. To plan just one prompt, prefix it with /plan.
3. Web or mobile — use the mode dropdown next to the prompt box at claude.ai/code and choose Plan.

To make it the default for a project’s terminal sessions, add this to .claude/settings.json in the project folder:

{
  "permissions": {
    "defaultMode": "plan"
  }
}

Worth knowing on desktop: the mode selector remembers most choices per folder, but Plan is the exception — picking it applies to that session only. New setup? Installing Claude Code covers the desktop route first.

How to read a plan when you can’t read code

Ask the plan five questions. All five are answerable in plain English:

• Does it build what I asked for? Count the features in the plan against the features in your request. Extras are the tell.
• What is it touching? Claude names the files it will change. Skim that list for surprises: a “change the button color” task has no business rewriting your login.
• Does it touch money or user data? Payments, passwords, email sending, deleting records. Those get a slower read than the rest.
• What’s missing? The plan is the last cheap moment to say “you forgot the confirmation email.”
• Does it explain itself? A plan you can’t follow is a plan you can’t approve. Ask Claude to rewrite it for a non-technical reader — that’s a fair request, and a good tool answers it.

A tighter starting brief produces a tighter plan. Our PRD template is the one-page version of that brief, and the copy-paste prompt templates cover the wording.

Approving, editing, or rejecting the plan

When the plan is ready, Claude presents it and asks how to proceed. You get three options:

• Yes, and use auto mode — approve and let a classifier review actions instead of you. When auto mode isn’t available on your account, this reads “Yes, auto-accept edits.”
• Yes, manually approve edits — approve the plan, then review each edit as it happens. Slower, and the right default while you’re learning what your tool does.
• No, keep planning — stay in plan mode and tell Claude what to change about the plan.

That third option is the one most people skip. Sending a plan back three times costs you five minutes; approving a wrong one costs you an afternoon of untangling.

You can also press Ctrl+G to open the proposed plan in your text editor and edit it directly before Claude proceeds — delete the step you don’t want, save, and Claude works from your version.

Approving a plan exits plan mode and switches the session into whichever permission mode you chose, so Claude starts editing right away. To plan again, cycle back with Shift+Tab or prefix your next prompt with /plan.

What trips people up

• Approving because the plan sounds smart. A confident plan for the wrong feature is still the wrong feature. Read it with your original request open beside it and match them line for line.
• Asking for the whole app in one plan. Plan one slice — one screen, one flow. A twenty-step plan is a twenty-step guess, and every step after a wrong one inherits the mistake.
• Forgetting it switched off. Approval ends plan mode. If you want the next task planned too, turn it back on.
• Treating the plan as your spec. The plan describes how Claude will build. Your PRD describes what you want and why. You still write that part.

The bottom line

Plan mode gives you one moment of real control, in the only place you can use it: before the code exists, while everything is still sentences you can read. Turn it on, read the plan, send it back when it’s wrong. That habit is most of the difference between founders who ship with AI and founders who get stuck in a codebase they can’t steer. The rest of the journey, from idea to product to income, is what Software Secrets 2.0 covers, and the whole book is free.