Prompt phrasing is subtle - what's the best practice?

In the Inferring section we look at a customer’s lamp review. I added these questions to the “Doing multiple tasks at once” prompt to see if I could tease out more details from the review.

  • Did the reviewer have problems?
  • Did the reviewer have difficulty with the product?
  • Did the reviewer encounter difficulty?

I’d expect “true” for all of them, since one part broke and another part was missing. But the answers were:

  • true
  • false
  • true

How should one phrase this kind of question? And how much validation needs to be done to ensure that the question leads to correct answers?

Note that ChatGPT is a language model, so there’s no guarantee it will ever give you a “correct answer”.

I get that, but if I’m contracted to write a chatbot for an ecommerce site, for instance, would they be happy with “well it’s fundamentally not an accurate tool”? I’m wondering how people get maximum quality and how they demonstrate that…

I guess the important thing would be to set expectations at the beginning of the conversation


Pretty much, they don’t. It’s a hype-driven market at the moment.