The Skip

A Claude Code confession

I started Claude Code like everyone does: carefully.

“Can I read this file?” Yes. “Can I run this command?” Yes. “Can I check your AWS account?” ...Yes? Every time, a prompt. Every time, a decision. Every time, I read what it was asking and made a judgment call.

I said yes to almost everything. In thousands of prompts, I almost never said no. But I paid attention. I learned what Claude Code does, how it thinks, what it asks for and why. I built a mental model of its behavior through repetition. That mental model would later become the justification for what I did next.

Mostly I was building simple websites. Side projects. Nothing mission-critical. A restaurant site here, a personal project there. And Claude Code was good. Surprisingly good. It understood what I wanted before I finished explaining it. It wrote clean code. It caught its own mistakes more often than not.

Good enough that the permission prompts started feeling like speed bumps on a road I'd already driven a hundred times. I knew the road. I knew the car. Why was it still asking me to confirm every turn?

That feeling — that impatience with safety you've outgrown — is the most dangerous feeling in technology. It's the feeling right before the mistake.

Then I found it: --dangerously-skip-permissions.

The name alone should have been a warning. It's not called --trust-mode or --auto-approve. It's called dangerously skip. Anthropic named it that on purpose. They wanted you to feel the weight of what you're doing.

I felt it. I typed it anyway.

$ claude --dangerously-skip-permissions

And just like that, the prompts stopped. The speed bumps disappeared. The road opened up and I was flying.

Since turning it on, I've moved faster. I've also made bigger mistakes.

Files modified that shouldn't have been. Configurations changed without me understanding why. Commands run that I would have questioned if I'd seen the prompt. Nothing catastrophic — these are side projects. But the pattern is clear: the more I skip, the less I understand what's happening. And the less I understand, the more I need to trust. It's a compounding loop, and it only goes one direction.

Each mistake is small. Each one is recoverable. But each one also moves the baseline. Yesterday's “I would never do that” becomes today's “it's probably fine.” The erosion of caution isn't dramatic. It's comfortable.

Anthropic didn't hide this flag. They didn't make it easy to find, either. And they named it with the word “dangerously” right there in the syntax. That's not an accident. That's a company that understands the moment a user crosses from informed consent to blind trust — and wants to make damn sure you know which side you're on.

Most companies would call it “expert mode” or “advanced settings.” Something that flatters the user. Something that makes bypassing safety feel like a promotion. Anthropic called it what it is: dangerous. And they put that word where you can't miss it.

I respect that. I also ignored it. And that's the whole problem, isn't it? You can name the danger. You can put it in bold. You can make people type it out letter by letter. And they'll still do it, because the speed is too good to give up.