I feel like we're talking about different things. You seem to be describing a mode of working that produces output that's good enough to warrant the token cost.

I feel like we're talking about different things. You seem to be describing a mode of working that produces output that's good enough to warrant the token cost. That's fine, and I have use cases where I do the same. My gripe was with the parent poster's quote: > Claude and GPT regularly write programs that are way better than what I would’ve written What you're describing doesn't sound "way better" than what you would have written by hand, except possibly in terms of the speed that it was written.