Claude Code's Plan Mode: Review the Plan Before It Touches Anything

The most expensive mistake an AI assistant can make isn't getting one step wrong. It's getting the direction wrong and then finishing all ten steps anyway.

You say "clean up this folder." It decides that means "sort everything into subfolders by file type" — and does it in five seconds, no hesitation. You're left staring at a structure you didn't ask for, reaching for Ctrl+Z. Except it moved real files. There's no Ctrl+Z.

Claude Code has a mode built specifically for this problem: Plan Mode. It splits "understanding the task" from "doing the task," and puts your approval in between.

What Plan Mode actually is

Plan Mode is a permission mode in Claude Code. With it on, Claude can read, inspect, and search — but it cannot change anything. What comes out the other end is a plan, waiting for your sign-off.

Allowed: reading files, searching content, listing directories, looking things up online. Blocked: writing or editing files, running any command that changes something.

The important part is that this isn't Claude being asked to behave. It's a hard restriction at the tool layer — write tools are blocked outright while the mode is on. That's a different thing entirely from telling it "don't touch anything yet." A reminder can be missed. A block can't.

Here's an everyday version of the same idea. You hire a contractor for a renovation. One version: he walks in and starts knocking down walls while you watch nervously. The other version: he walks the site first, then hands you a page — "here's what I'd do: this wall comes down first, wiring runs this way, three days total." You read it, nod, and only then does work start.

Plan Mode is the second contractor. And the payoff shows up exactly when step two of the plan turns out to be wrong — that's the moment you just saved yourself from redoing the entire job.

How to turn it on

Three ways in, plus one setting that makes it stick.

1. Keyboard toggle (the one you'll use most)

Press Shift+Tab during a Claude Code conversation. It cycles through several permission modes, and plan mode is one of them. Once active, the interface shows a plan mode on status indicator.

Press Shift+Tab again to keep cycling out of it, and the plan never gets executed.

2. One-off use: /plan

If you only want a plan for this single instruction, without changing the mode for the rest of the session, prefix the command:

/plan reorganize this folder by month

That one turn runs in Plan Mode and hands you a plan to approve.

3. Launch straight into it: --permission-mode

Set it at startup:

claude --permission-mode plan

Run claude --help to see the full list of accepted values — currently six: acceptEdits, auto, bypassPermissions, manual, dontAsk, plan.

4. Make it the default

If you want every session to start in plan mode, set it in settings.json:

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

In ~/.claude/settings.json, this applies machine-wide. In a project's .claude/settings.json, it applies only to that project — and the project-level setting wins over the machine-level one. Commit that line into a shared project and you've added a default safety brake for the whole team.

Once you have a plan: four ways to respond

A plan from Claude isn't a simple approve-or-reject. You actually have four options:

The third option is the most underrated one. The real value of Plan Mode isn't "stopping Claude" — it's that a plan is text, and editing text is cheaper than undoing a result. Spot a wrong step two, say "leave the config alone in step two, just rename the files," and it's fixed. Catch the same mistake after execution, and the fix is redoing the whole thing.

When it's worth using — and when it isn't

Plan Mode isn't something to leave on for everything. It has a cost: one more round-trip, a bit more waiting.

Turn it on when:

1. The action can't be undone. Bulk renaming, moving or deleting files, overwriting existing content. Get it wrong and there's no undo.

2. The scope is fuzzy. "Clean up this pile of files" — if you haven't fully worked out the target structure yourself, having Claude propose one first often clarifies your own thinking.

3. Multiple files are involved. Checking scattered changes after the fact takes far more effort than reviewing one plan up front.

4. You're in unfamiliar territory. Having it explain the approach first means you pick up how the task should be done, not just get it done.

Skip it when:

• It's pure read-and-answer ("what does this file say") — nothing was going to change anyway.

• The change is small, specific, and reversible ("change the date format in this file to YYYY-MM-DD").

• Your instruction was already precise enough that the plan would just repeat it back to you.

One rule of thumb: if getting it wrong would cost you more than five minutes to fix, look at the plan first.

A non-coding example

Plan Mode gets filed under "developer feature," but it isn't one. Here's a paperwork example instead.

You've got a Downloads folder that's been accumulating for two years — hundreds of files, no naming convention. You want it sorted.

Without Plan Mode, you say "clean up my Downloads folder" and it might just start moving things — sorted by file type, with invoices and contracts dumped into the same "documents" folder, when what you actually wanted was sorting by year and client.

With Plan Mode on, you get a plan first: scan every file and tally up types and dates, propose a folder structure, flag anything it can't classify, then wait for your go-ahead before moving a single file.

You read it and say: "not by type — by year. Invoices get their own folder. Anything without a clear date goes into unsorted, don't guess."

One sentence, and you've just saved yourself an afternoon of moving hundreds of files back. Same logic applies to summarizing a stack of PDFs, merging scattered meeting notes into one document, or batch-editing headers and footers across files. All irreversible, all worth a plan first.

Two things worth knowing

First: Claude can still search the web while in Plan Mode. Web search and page reads are read-only actions, so they're unrestricted. That's a good thing — the plan gets built on verified information, not guesswork.

Second: for complex tasks, Claude may spin off separate helper agents to do the reading during the planning phase. Those helpers are read-only too — the point is to offload heavy lookup work into a separate context rather than filling up your main conversation. Seeing it "go off and check something, then come back" is expected behavior, not a glitch.

This connects to a related idea worth reading if you haven't: our earlier piece on context management, "Why Claude Code Gets Expensive — and the Three Commands That Fix It" https://essevin.com/blog/claude-code-context-compact-clear-en?inv=zd9km4gn&lang=en . Plan Mode and context management are really two sides of the same coin — one cuts down on mistakes, the other cuts down on paying twice for the same work.

The step before any of this

Everything above assumes Claude Code actually runs for you. Anthropic officially supports 175 countries and regions, and neither Hong Kong nor China is on that list — direct sign-up doesn't work.

Essevin's approach: two environment variables connect Claude Code to a stable gateway — direct connection, no code changes, no need for a foreign credit card. One account gives you access to multiple top-tier models, billed by usage. Top up with FPS or PayMe; users in China can pay with Alipay or WeChat Pay just as easily. Full setup steps: "Claude Code Setup in China: Complete Guide — from Registration to First Autocompletion in 10 Minutes."

Once you're connected, the first thing worth doing is pressing Shift+Tab to switch into Plan Mode and pointing it at a folder you've been avoiding for years. Reading the plan it hands back will teach you more about this mode than any write-up could.

Information in this article is current as of 25 July 2026 and is provided for general reference only; it does not constitute advice of any kind. Third-party product features, pricing and policies are subject to their official announcements. Essevin service details are as shown on essevin.com and in the console.