Either it depends a lot on the model, or I’m bad at chat, or both, because in my experience once a chat starts going off the rails it’s pretty hard to correct the initial misconception. I find it works better to backtrack to just before the misconception and edit my previous response to head it off at the pass.

Definitely agreed that it's better to not have some text in the context than to have some that needs to be opposed. I suppose I meant steering prompts are a non-obvious innovation to LLM interaction.