Most of what you speak of is caused by this simple statement from the onset:

You use it as a tool, that’s it.

If you use it as a tool, then expecting anything but minimal compliance would be an irrational expectation, because it will do exactly as you ask, which is the bare minimum to satisfy your requirements.

I got tired of that, so as an experiment, I decided to go about it differently, and made a couple discoveries about how to make it work. Turns out LLMs can’t pattern match things they don’t have in their training sets.