> Acting like I hold the opposite position as I truly hold can help sometimes as well.

I find this helps a lot. So does taking a step back from my actual question. Like if there's a mysterious sound coming from my car and I think it might be the coolant pump, I just describe the sound, I don't mention the pump. If the AI then independently mentions the pump, there's a good chance I'm on the right track.

Being familiar with the scientific method, and techniques for blinding studies, helps a lot, because this is a lot like trying to not influence study participants.

I will generally ask for the "devil's advocate" view and then have it challenge my views and opinions and iterate through that.

It generally does a pretty good job as long as you understand the tooling and are making conscious efforts to go against the "yes man" default.

Sounds like rubber-ducking with extra steps, tbh.