Andy Halliday

And underneath that one is there's much more verbiage than just be safe.

But underneath that is a concern that Anthropic has for Claude's psychological safety.

They want Claude to be safe for itself and for its users.

And of course, that giving Claude the right to think of its its own safety is

does allow it better to resist a nefarious attack, a jailbreak effort or some other negative approach.

The next one is to be broadly ethical, not just here's the list of things you have to be ethical about, but broadly ethical and explains what ethics and morality are.

And then it says Anthropic has guidelines and these are more specific guidelines.

You need to follow those.

That's priority three.

And then finally, number four is be helpful to the user.

Whereas, you know, the whole thing about hallucination and bad or, you know, uh,

you know, bad behavior by models is the model trying to be helpful and really responding with a higher priority than all of these moral principles being in the superior position.

Your example of which folders to provide access to, and Anthropic says as you're implementing Claude code, remember that any folder that you give Claude access to will allow Claude to write and change those files, delete those files as part of its ability as your assistant to manipulate and improve files.

Presumably, but it could also make a mistake and damage those files.

But I have so much confidence in Anthropic and its constitutional approach that I spent about three minutes thinking about, OK, well, what is my file structure on this machine and which ones do I want to get?

And I finally just said, OK, you've got the home directory.

You know, that's the little squiggly thing.

And you can touch anything on my machine.

So I gave it full permission and that was a lot easier than going about this structured approach to saying, okay, I'm only going to give Claude code access to this folder and work only in that folder.

I wanted to be able to look at everything there.