The basic insight is that you can iterate like this without losing coherence and starting to generate artifacts from feedback loops and overfitting.

I also figured out you get much better responses if you ask it to write prose responses instead of bulleted lists. It's kind of navigating semantic latent space so if you ask it to respond in different styles (like prose + free verse + anglo saxon poetry style) I think it helps it cover a lot more ground.

Asking it about qualia+feelings+thoughts about the 10 answers in a debrief afterwards seems to be a good procedure too.

I think treating it respectfully does have a big objective impact in the prompt result.