>>105673579 (OP)i found a new balance.
Now when i start a new environment/language/APIs i let the AI do as much as it can. Write all the base functions/classes, solve all the basic issues, give a try to the complex ones. Always scold it to encapsulate and simplify everything it can, use the simplest types and "are you sure that's the best you can do?" prompts.
Once i'm down 20k lines, and issues crop up i have to solve about everywhere, i actually begin doing slower work. I rewrite what looks jury rigged. I simplify things further. Unavoidably i see the dumb indian-tier solutions mixed with textbook ones. I rewrite or ask it to do things differently. By the time i'm done, i was still 5-6 times faster but know the codebase and have written enough in it to have a good idea of the APIs, syntax, language quirks. Then i can study or wrap up the project, or have high level thinking about refactoring.
Big gain: time, efficiency, not getting boggled in details, VERY fast learning the more i force myself to write the code
Loss: actually lowered skill since i'm not solving low order issues myself, and get the feeling i'm solving more than i actually am
The tool is actually super-useful, but the human brain shows how obsessed it is with conserving thinking power. I have to force myself to learn the codebase, so the loss of skill is compensated by the amount of learning i do.