How to improve prompts for davinci-instruct

I’m creating this thread as I’d like to know what others have found useful in prompt creation. If you could share your insight as to what has been working and what hasn’t been working. For what hasn’t been working add solutions to others if you know what could be helpful.

I’ll go first.

Good: I find that my prompts have been working well when it is a defined point that can easily be found. For example, ‘include the history of x’ or ‘include the location of x’.

Bad: It is really hard at pulling data or numbers that I am requesting. ex: It should include data from 2021. It usually just pulls any data available sometimes from a decade ago. Any solutions on this?

1 Like

How are you able to integrate a corpus or dataset where it will only pull one aspect of the prompt. My sense is if you have a corpus or dataset it will use that as the main source for all aspects of the prompt.

Also what have you found to work well or not work well?

@llimateuslima curious if you have anything to share as well

IMO you should never be using GPT to try to get concrete values directly. You use it to get “common sense” or to provide a wrapper/translator/extractor over some sort of traditional search backend.
You could use a prompt like:

If the following message contains a question, then format the question as a google search that is likely to provide a useful answer. 

Ah! I like that. Is it possible to use a prompt like this: