Permissions: Stop Clicking 'Allow' on Everything

The Permission Problem

You ask Claude Code to refactor a component. It wants to read a file — allow. Edit a file — allow. Read another file — allow. Run a test — allow. Five clicks before anything happens.

This is the number two complaint about Claude Code (number one is cost). And it's completely fixable. You just have to configure it once.

Setting Up allowedTools

The magic lives in your settings.json. There are three levels:

Project-level (.claude/settings.json in your repo) — shared with your team via git. Best for project-specific tools.

User-level (~/.claude/settings.json) — your personal defaults across all projects.

Here's what a solid starting config looks like:

{
  "permissions": {
    "allow": [
      "Read",
      "Glob",
      "Grep",
      "LS",
      "Bash(npm test*)",
      "Bash(npm run lint*)",
      "Bash(npx tsc*)",
      "Bash(git status*)",
      "Bash(git diff*)",
      "Bash(git log*)"
    ],
    "deny": [
      "Bash(rm -rf*)",
      "Bash(git push --force*)",
      "Bash(git reset --hard*)"
    ]
  }
}

Read operations are always safe to auto-accept. Test and lint commands are safe. Destructive git commands? Always review.

The /permissions Command

Already in a session and tired of clicking? Type /permissions to manage permissions on the fly.

Trust Levels: What to Auto-Accept vs Always Review

Always auto-accept (safe):

• Read — reading files never changes anything
• Glob — searching file names is harmless
• Grep — searching file contents is harmless
• LS — listing directories is harmless
• Bash(git status*) — read-only git commands
• Bash(git diff*) — read-only git commands
• Bash(git log*) — read-only git commands

Auto-accept with patterns (mostly safe):

• Bash(npm test*) — running tests
• Bash(npm run lint*) — running linters
• Bash(npx tsc*) — type checking
• Bash(cat *) — reading file contents
• Edit — if you trust Claude Code's edits (most people get here after a week)

Always review (potentially destructive):

• Bash(rm *) — deleting files
• Bash(git push*) — pushing to remote
• Bash(git checkout*) — switching branches
• Bash(git reset*) — resetting changes
• Bash(npm publish*) — publishing packages
• Bash(curl*) — making network requests
• Any command that touches production

The Progressive Trust Approach

Don't configure everything on day one. Start conservative and open up as you build trust:

Week 1: Auto-accept reads only. Review everything else. Learn what Claude Code actually does.

Week 2: Auto-accept reads + test/lint commands. You now trust it won't run weird shell commands.

Week 3: Auto-accept edits too. You've seen enough edits to trust the quality.

Week 4+: You're only reviewing destructive operations and network calls. That's the sweet spot.

One More Thing: The Accept Button

When a permission prompt does appear, you have options beyond just "allow":

• y — Allow this once
• a — Allow this tool for the entire session
• n — Deny this request
• Escape — Cancel the current operation

The "a" option is your friend for exploratory sessions where you trust the general direction but didn't pre-configure permissions.