By default Claude pauses and asks before it changes a file or runs a command. In Claude Desktop this shows up as a clear approve/decline prompt right in the chat — the docs call the gentle version the "Ask permissions" mode and recommend it for new users. As you build trust, you can let the boring, safe actions run on their own while Claude still checks in on anything risky. You're always in control of where that line sits. (In the terminal, the /permissions command is the Power-Track way to manage the same rules.)
Out of the box, Claude pauses before it changes a file or runs a command and asks you first. In Claude Desktop this is a simple Approve/Decline prompt right in the chat — the docs call the gentle, asks-about-everything-risky version the “Ask permissions” mode and recommend it for new users. It’s the right place to start: nothing happens behind your back, and you learn the tool with the brakes on.
The instinct to follow is tighten before you loosen. Start cautious, let the safe, boring actions earn the right to run on their own one at a time, and you end up answering exactly the right question — “is this destructive?” — at exactly the right moment. Reads and tests can flow freely; deletes, force-pushes, and anything irreversible stay behind a human yes.
Permissions are project-scoped, which is the quietly important part: a project you
barely trust can stay strict while a sandbox you’re playing in runs wide open, and
the two never bleed into each other. On Desktop you set this from the mode selector;
in the terminal, the Power-Track /permissions command manages the same rules.
The point isn’t fewer guardrails — it’s putting the guardrails only where they matter.
Generator
Build a settings.json
Pick permission rules and hooks, then copy out a settings.json for your .claude/ folder.
Permissions
allow run without asking
ask always confirm first
deny never, full stop
Hooks
{
"permissions": {
"allow": [
"Bash(npm run test:*)"
],
"deny": [
"Read(./.env)",
"Bash(rm -rf:*)"
]
},
"hooks": {
"PostToolUse": [
{
"matcher": "Edit|Write",
"hooks": [
{
"type": "command",
"command": "npx prettier --write ."
}
]
}
]
}
}
A starting point — copy it into your project and adapt.
why it helps Stop clicking 'approve' on harmless reads, without losing the guardrails that matter.
examples- On Desktop, start in the "Ask permissions" mode (the one to start in) — it asks before risky actions and keeps you in control.
- Let reads and tests run freely; keep a human in the loop for anything destructive.
- Tighten before you loosen — start cautious, then relax the rules as you build trust.
- Permissions are project-scoped, so a risky project can stay strict while a trusted one runs free.
Questions people ask
- Does Claude ask before it changes my files?
- Yes, by default. In Claude Desktop you get a clear approve/decline prompt in the chat before Claude changes a file or runs a command — the recommended "Ask permissions" mode for new users. As you build trust you can let the boring, safe actions run on their own while Claude still checks on anything risky.
- What is the \"Ask permissions\" mode?
- It's the Claude Desktop permission setting the docs recommend for new users: Claude asks for your approval before it does anything that changes your files or runs a command, with a simple Approve/Decline prompt. You stay in control and learn the tool without anything happening behind your back.
- Are permissions shared across all my projects?
- No — permissions are project-scoped. A project you barely trust can stay strict while a sandbox you're playing in runs wide open, and the two never bleed into each other.
- What should I approve freely and what should I keep asking about?
- Let reads and tests run freely; keep a human in the loop for anything destructive — deletes, force-pushes, anything irreversible. Tighten before you loosen: start cautious, then relax as you build trust.