AI Always Asking "May I Run This?" Cut Prompts with /permissions

Using Claude Code, have you had endless confirmation prompts like "may I run git status?" or "may I read this file?" It's for safety, but asking every time even for obviously safe commands keeps breaking your workflow. Yet allowing everything makes you anxious that dangerous commands like delete or deploy might run on their own. /permissions strikes the balance between the two. You set rules once: allow the safe, block the dangerous.

Definition

/permissions is a command to set rules for how far Claude may run without permission.

The rules split broadly into three kinds.

• allow: Commands that may run right away without asking. (e.g., reading a file, checking changes)
• ask: Commands that ask once before running. (the default)
• deny: Commands that are never run. (e.g., deleting a file)

Put frequently-used safe tasks in "allow" and "may I?" prompts drop sharply, and block dangerous tasks with "deny" to prevent accidents. In short, it's a "permission traffic light" you give the AI.

How to use it (by difficulty)

Basics — open the rules window

Type into the chat box.

/permissions

On entering it, the currently set permission rules (allow/ask/deny lists) appear on screen. Here you can add or remove rules. At first it'll be mostly empty, and you fill it in with the methods below.

Applied — allow safe read commands

Commands like git log (view commit history) or git diff (view changes) only read and change nothing, so there's no need to ask every time. Just ask in conversation like this.

from now on, run git log and git diff right away without asking

Then Claude adds these commands to the allow list. From then on they run right away without a confirmation prompt.

Advanced — block only dangerous commands, pinpoint

Conversely, you can block dangerous commands entirely.

block file-delete commands (rm) so they're always denied

Set this and delete-type commands won't run, by mistake or otherwise. It captures safety and speed at once, like "read freely, block deletion at the source."

Common pitfall — "permission mode" and "permission rules" are different

This is where beginners get confused. Distinguish two things.

• Permission rules: What you set with /permissions. The detailed rules deciding "which commands to allow/deny."
• Permission mode: The overall attitude. A big mode like "ask every time, or proceed on its own." This is switched not with /permissions but by pressing the Shift+Tab key in the chat box. (Each press cycles the mode.)

So "it asks too often" can be solved two ways. Put individual commands in the allow list (/permissions), or change the mode itself (Shift+Tab). Don't confuse the two.

A real case

It kept asking "may I run git status?", "may I look at this file?" every time I worked, breaking my flow. So I put the read commands I use often into the allow list all at once.

allow read commands like git status, git log, git diff without asking

After that, confirmation prompts almost disappeared. With reading code and checking changes flowing smoothly, my perceived speed doubled. Instead, I left dangerous commands like delete and deploy still asking, so safety stayed intact.

Use it like this too

• Allow safe commands: Put frequently-used read commands in the allow list to cut prompts.
• Block dangerous commands: Keep hard-to-undo commands like delete and deploy always on ask or deny.
• Switch modes: To change the big flow, switch permission mode with Shift+Tab in the chat box.
• Step by step: Don't allow everything from the start — add to the allow list one by one, starting with what feels "this is safe."

Tip: It's safest to fill the allow list starting with "read-only commands." Leave commands that change or delete files on ask if possible. One mistake leading to a big accident is mostly on the "write/delete" side.

Wrap-up

The key is one sentence. Allow the safe, block the dangerous. Get worn down by constant prompts and allow everything, and accident risk grows; block everything and it's stifling. Set rules once with /permissions and you secure safety and speed at the same time.

Based on: Claude Code v2.1.154 (2026.05)