For the chat systems I've found acting like Columbo (from 1970's TV detective) works wonders: you want to be polite but persistent, open but not gullible. Don't fight it, but don't just let it drive.
For the non-chat interfaces, I imagine a whiteboarding session with a really competent intern at the board, rapid prototyping / wireframing that you can play with "live" and refine far further than you could IRL, but still ultimately prototyping.
> I've been continually surprised by the number (and variety) of people who don't see the utility of it.
If you _don't_ do it this way, you can easily fall into all sorts of time wasting anti-patterns; if you try to trick it, or allow yourself to be easily fooled by it, get stubborn & closed minded, pedantic and argumentative or whatever, well, there are lots of examples of how those sorts of interactions go in the training data too, and it will just as happily go down them as any other.
