Claude Code Dangerously Skip Permissions: a Sane Guide

The flag --dangerously-skip-permissions tells Claude Code to stop asking and just do: edits, commands, network calls all run without prompts. It's one of six permission modes, and the word 'dangerously' is in the name on purpose — there are calmer ways to get fewer interruptions.

I'm Evgeny, and I'll admit it: every newcomer eventually googles this flag, usually around prompt number forty in a single evening. So instead of pretending you won't, let's understand the whole permission system — what each mode trades away, and which setup gives you speed without handing Claude the keys to your laptop.

What does --dangerously-skip-permissions actually do?

It starts Claude Code in bypassPermissions mode: permission prompts and safety checks are off, and tool calls execute immediately. By default Claude pauses before editing files, running shell commands, or touching the network; this flag removes those pauses. Two things still stop it: explicit 'ask' rules you've configured, and a circuit breaker for catastrophic deletions — commands like rm -rf / or rm -rf ~ still prompt even here. Honest translation for non-programmers: the flag means 'don't ask me anything', and you should only say that inside an isolated container or VM where a mistake can't hurt your real machine.

claude --dangerously-skip-permissions

The six permission modes at a glance

Claude Code has six modes, each a different trade between convenience and oversight. default: only reads happen without asking — best for starting out. acceptEdits: file edits and common filesystem commands like mkdir, touch, mv, cp run freely — good while you review changes via git diff afterwards. plan: read-only research mode, Claude proposes before touching anything. auto: everything runs, but a separate safety classifier reviews each action in the background. dontAsk: only tools you pre-approved run, everything else is auto-denied — built for scripts and CI. bypassPermissions: everything runs, no checks — containers and VMs only.

In every mode except bypassPermissions, writes to protected paths — things like .git and Claude's own .claude config folder — are never auto-approved. That quietly protects you from the classic 'the AI edited its own permissions' horror story.

How do you switch modes — is there a shortcut?

Yes: Shift+Tab cycles default → acceptEdits → plan mid-session, and the status bar shows where you are. bypassPermissions only appears in the cycle if you started with an enabling flag — you cannot wander into it by accident. There's even a polite variant: --allow-dangerously-skip-permissions adds the mode to your Shift+Tab cycle without activating it, so it's one keypress away when a trusted task needs it.

Can you skip permissions by default?

You can set any default mode in settings — but please pick a sane one. The defaultMode setting in .claude/settings.json controls how every session starts. My recommendation for non-programmers after the first week: acceptEdits as default, plan mode for big tasks, and bypass only inside disposable environments.

{
  "permissions": {
    "defaultMode": "acceptEdits"
  }
}

Safer alternatives: auto mode and friends

If your real complaint is prompt fatigue, the modern alternative to skip-permissions is auto mode: Claude works without routine prompts while a separate classifier model watches every action and blocks anything that escalates beyond your request — production deploys, force pushes, curl-pipe-bash downloads. It's a research preview, not a guarantee, but it's the 'fewer prompts without removing the brakes' option. The second lever is fine-grained rules: run /permissions and add allow rules for commands you trust, like Bash(npm run *), so the routine stuff stops asking while everything else still does.

Why is the flag 'not working' for some people?

Three common cases. You're root or using sudo — blocked by design on Linux and macOS. You tried to switch into bypassPermissions mid-session without having started with an enabling flag — restart with the flag. Or an administrator set disableBypassPermissionsMode in managed settings, in which case the decision was made above your pay grade, and honestly, probably correctly.

Kicker: permissions aren't bureaucracy, they're a dial. Start tight, loosen exactly where it annoys you, and save 'dangerously' for places where danger physically can't reach. Tonight, instead of the big red flag, just add three allow rules — you'll feel the difference by prompt ten.