By default, Claude Code asks before it runs a command or edits a file. That's safe, but the prompts add up. The /permissions command lets you decide in advance what's allowed and what's off limits.
Opening the rules
Run /permissions to view and edit your allow and deny lists:
/permissions
Rules are written per tool. You can allow or deny broad tools like Read, Edit, and Write, or narrow patterns for shell commands and paths:
Bash(git *)
Bash(npm run test:*)
Write(/tmp/*)
Edit(src/*)
An allow rule means Claude does that thing without asking. A deny rule blocks it outright. So you might allow Bash(git *) because you trust git commands, while denying writes outside your project.
Permission modes
Alongside rules, Claude runs in a mode that sets the overall tone:
- default — ask before acting.
- acceptEdits — auto-accept file edits, still ask for commands.
- plan — read-only; Claude looks but changes nothing.
- bypassPermissions — no prompts at all. Powerful and risky; use with care.
Cycle through modes quickly with Shift+Tab without opening any menu.
A sensible setup
Start strict and loosen as you build trust. Allow the safe, repetitive things you approve every time — your test runner, git, reads — so the prompts stop interrupting you. Keep denies for anything destructive or outside your project. The goal is fewer interruptions on the boring stuff while keeping a real gate on the dangerous stuff.
Comments
Be the first to comment.