In general, I tend to do better when people point out what I got wrong, rather than assuming I'm trying to be adversarial(?) when I'm not really. I do have more specific instructions for workflows that the LLM doesn't seem to have pre-existing knowledge of, but I don't over-explain things that it can clearly already do on its own. Aka: I get more specific when it screws up, but I don't over-explain when it doesn't. I start with basic facts, and refine from there if it doesn't get the idea.