DEV Community

xujfcn
xujfcn

Posted on

How to Skip Permission Prompts in Claude Code: The --dangerously-skip-permissions Flag Explained

#tutorial #cli #ai #productivity

If you've used Claude Code (Anthropic's CLI for Claude), you know the pain: every file edit, every shell command, every tool call triggers a permission prompt.

Claude wants to edit src/index.ts
Allow? (y/n)
Enter fullscreen mode Exit fullscreen mode

For a 10-file refactor, that's 30+ prompts. It kills your flow.

The --dangerously-skip-permissions flag exists for exactly this reason. But the name is scary on purpose. Let's break down what it actually does, when to use it, and how to stay safe.

What Does --dangerously-skip-permissions Do?

By default, Claude Code asks for your approval before:

  • Editing or creating files
  • Running shell commands
  • Accessing external tools (web search, APIs, etc.)

The --dangerously-skip-permissions flag auto-approves all of these. Claude will read, write, execute, and modify without asking.

claude --dangerously-skip-permissions
Enter fullscreen mode Exit fullscreen mode

That's it. No more prompts.

Why It's Called "Dangerously"

Anthropic named it this way to make you think twice. Without permission prompts, Claude can:

  • Overwrite files without confirmation
  • Run shell commands including destructive ones (rm -rf, git push --force)
  • Install packages you didn't ask for
  • Modify system configs if it has access

In a production environment or a repo with uncommitted changes, this could be catastrophic.

When It's Actually Safe to Use

Here's the thing: for most development workflows, the risk is low if you follow basic precautions.

Safe scenarios:

  • ✅ Working in a git repo with clean commits (you can always git reset)
  • ✅ Greenfield projects (nothing to break)
  • ✅ Disposable environments (Docker containers, VMs, CI/CD)
  • ✅ Code review tasks (Claude is reading, not writing)
  • ✅ Automated pipelines where you review output afterward

Dangerous scenarios:

  • ❌ Production servers
  • ❌ Repos with uncommitted work
  • ❌ Directories with sensitive files (SSH keys, env files, credentials)
  • ❌ Running as root without sandboxing

The Smart Way: Use It with Safeguards

1. Always commit before starting 

git add -A && git commit -m "checkpoint before claude session"
claude --dangerously-skip-permissions
Enter fullscreen mode Exit fullscreen mode

Now if anything goes wrong: git reset --hard HEAD

2. Use in a specific directory 

cd ~/projects/my-safe-project
claude --dangerously-skip-permissions
Enter fullscreen mode Exit fullscreen mode

Don't run it from your home directory or root.

3. Combine with a task description 

claude --dangerously-skip-permissions -p "Refactor the auth module to use JWT tokens. Don't touch the database layer."
Enter fullscreen mode Exit fullscreen mode

Clear instructions reduce the chance of Claude going off-script.

4. Use Docker for full isolation 

docker run -it -v $(pwd):/workspace -w /workspace \
  node:22 bash -c "npm i -g @anthropic-ai/claude-code && claude --dangerously-skip-permissions"
Enter fullscreen mode Exit fullscreen mode

Even if Claude runs rm -rf /, it only affects the container.

Alternative: Permission Allow Lists

If --dangerously-skip-permissions feels too aggressive, Claude Code also supports allow lists in your project settings. Create a .claude/settings.json:

{
  "permissions": {
    "allow": [
      "Edit(src/**)",
      "Edit(tests/**)",
      "Bash(npm test)",
      "Bash(npm run build)"
    ],
    "deny": [
      "Edit(.env*)",
      "Edit(*.key)",
      "Bash(rm *)",
      "Bash(git push*)"
    ]
  }
}
Enter fullscreen mode Exit fullscreen mode

This gives Claude freedom to edit source code and run tests, but blocks it from touching secrets or pushing to git. Best of both worlds.

Using It in Automated Pipelines

The flag really shines in CI/CD and automation. If you're using Claude Code for automated code review, refactoring, or test generation:

# In a GitHub Action or CI pipeline
claude --dangerously-skip-permissions -p "Review the PR diff and suggest improvements" --output-format json
Enter fullscreen mode Exit fullscreen mode

In automated contexts, there's no human to click "Allow," so the flag is essentially required.

Pro Tips

Pair with an API gateway for cost control

When Claude runs autonomously, it can burn through tokens fast. Using an API gateway with budget limits prevents surprise bills:

# Use a gateway with spending caps
export ANTHROPIC_BASE_URL="https://crazyrouter.com/v1"
export ANTHROPIC_API_KEY="your-gateway-key"
claude --dangerously-skip-permissions
Enter fullscreen mode Exit fullscreen mode

Most gateways let you set daily/monthly spending limits, so even if Claude goes wild, your wallet is safe.

Check what Claude did afterward 

# After a session, review all changes
git diff
git diff --stat

# Or use git log if Claude made commits
git log --oneline -10
Enter fullscreen mode Exit fullscreen mode

Set up a pre-commit hook as a safety net 

# .git/hooks/pre-commit
#!/bin/bash
# Prevent accidental commits of sensitive files
if git diff --cached --name-only | grep -E '\.(env|key|pem|secret)'; then
  echo "⚠️  Sensitive file detected in commit. Aborting."
  exit 1
fi
Enter fullscreen mode Exit fullscreen mode

TL;DR

Approach Safety Convenience Best For
Default (prompts) ⭐⭐⭐⭐⭐ ⭐⭐ Production, sensitive repos
Allow lists ⭐⭐⭐⭐ ⭐⭐⭐⭐ Daily development
--dangerously-skip-permissions ⭐⭐ ⭐⭐⭐⭐⭐ Greenfield, CI/CD, Docker
Skip + git checkpoint ⭐⭐⭐⭐ ⭐⭐⭐⭐⭐ Recommended for most devs

My recommendation: git commit → --dangerously-skip-permissions → review changes → commit or reset. Fast, safe, reversible.

How do you handle Claude Code permissions? Share your workflow in the comments.

Top comments (0)