Ah that I will admit. It gets shit done one way or another haha. This is why a sandboxed environment and a reproducible test DB is key here. I give read only access to my dev DB to my Claude, really removes the temptation that it increasingly has to “cheat”. E.g. doing something hacky and fixing the DB manually in a way that doesn’t solve the problem everywhere.

Personally I love when the AI has this amount of problem solving. But you have to build the environment around it that encourages solving problems right the first time, versus taking the easy way out and hacking out a solution.

It’s just all about constraining the behavior of the LLM into productive and permanent directions. The more advanced it gets, the more it feels like designing engineering processes rather than coding. Personally it’s a fun change of pace and it’s giving me a lot of opportunities to look at the project in working on at a wider lens. I find having to pump out features makes you myopic in a sense. I really miss the control I had over writing it all by hand, but I love just being able to build software. At the end of the day, what do you want? That’s the question I’ve had to grapple recently.

Personally I don’t mind switching gears to the bigger picture of why the software exists and what purpose it serves

This honestly sounds like a tweaked system prompt more than anything. Maybe it is an attempt to make the model appear stronger?