Dangerously Skip Permissions Is Worth It

So my friend Spencer has been getting used to Claude Code, and he hit the wall everyone hits in the first hour. Every tool call, Claude asks: can I run this? Can I edit this file? Can I create this folder? You click yes, yes, yes, yes. After a while it feels like you are babysitting the thing instead of working with it.

There is a flag for that. It is called dangerously skip permissions. You launch Claude Code with it like this:

claude --dangerously-skip-permissions

And Claude stops asking. It just does the work.

Okay But Why Is It Called "Dangerously"

Because Anthropic is covering themselves. Claude Code is still in a fast-moving place. Features ship weekly. Things are labeled alpha, things are labeled beta, things change.

Quick detour on that word. If you are not familiar with it, beta basically means: we are putting this out there, we know it might not be perfect, but we need real people banging on it to figure out what breaks. It is a version that works well enough to use, not well enough to be called finished. Gmail was in beta for what felt like forever. There is a running joke that some products are just permanently in beta, and that is probably even more true with AI right now given how fast everything is changing. When you are on the bleeding edge, there might be a little bit of blood.

When you give a tool permission to write files, delete files, run shell commands, and send messages without asking first, there is always going to be a warning label on the box. That is appropriate.

But here is how I think about it. There is a risk in crossing the road. There is a risk in waking up in the morning and driving to the grocery store. You do it anyway, because the cost of staying home is higher than the small chance something goes wrong.

Same thing here. I went about three minutes with permissions on before I was like, give me the dangerously skip permissions flag. I have been running it for the last three months. I do not regret it for a second.

What Can Actually Go Wrong

Let me be honest about this. Things have gone wrong.

Claude has occasionally been a little aggressive. It has changed code I did not want changed. It has moved files I did not want moved. A couple of times it sent a text message or an email to a real person that I wish it had not sent.

Okay so those things happened. Here is why I still run the flag.

Code changes are reversible. If you are keeping your code in GitHub, which you should be, anything Claude writes can be undone with a git revert. You look at the diff, you see what it did, and if you do not like it, you roll it back. About a thirty-second fix.

File deletions are mostly reversible. You have iCloud backups, you have Time Machine, you have Google Drive history. Very few things in a modern setup are absolutely, permanently gone.

The things that are not reversible, you learn to gate. Sending a text to a real person. Sending an email. Publishing something live. Those are the ones where I put a permission back in. In my CLAUDE.md file, I have a rule that says: "Never send a message to a real person without showing the exact text first and getting explicit approval." Claude reads that rule every session. It asks before it sends. That is the compromise, skip permissions everywhere except the places where the action leaves the building.

What It Actually Feels Like

Before the flag, every session has this rhythm of stop, click, stop, click. You are paying attention to the approval prompts, not the work.

After the flag, Claude just works. You describe what you want, it goes and does it, and you look at the result. You review the diff, you test the change, you move on. It is closer to having a junior dev sitting next to you than a toddler who asks permission for every crayon.

The time difference is super real. A task that used to take fifteen minutes because I was clicking yes every thirty seconds now takes five. Right?

How to Actually Use It

Launch Claude Code with the flag:

claude --dangerously-skip-permissions

Put your gates in your CLAUDE.md file at the project level or in your global one. Things like:

• Never send a message to a real person without approval
• Never push to production without approval
• Always show the draft before publishing

Claude reads those every session. The permissions come back exactly where you want them.

Would I Recommend It

Yeah. Three months in, no regrets.

If you are new to Claude Code, spend the first hour or two with permissions on so you get a feel for what Claude is doing. Watch it work. Get comfortable with how it thinks. Then flip the flag and never look back. Your first day with Claude Code walks through the setup from scratch if you are not there yet.

If you have been clicking yes every thirty seconds for a while now, and you have a CLAUDE.md with your gates in it, just turn it on. The flag is worth the warning label.