I find the same thing when a compiler's fast and has good error messages. I can type vaguely what I want, and then the errors teach me the language syntax. (When there are small, clear examples, I can learn a lot from that as well. I really like Python IDLE.)

It sounds like most of the benefit you got from ChatGPT lines up with deficits in your tooling. It's interesting that a language model set to "predict" can make up for that deficit for such a wide variety of languages! (Though it's no good for niche or more-than-slightly technical scenarios, like Vulkan programming; past a certain point, you need the actual tools to exist.)