You want a base model like text-davinci-001. Instruct models have most of their creativity destroyed.
How do you use the base model?
OpenAI Playground still has it. Otherwise go out and find one.
I mean in terms of prompting. What methods do you employ to get useful results out of a model that is not tuned for particular form of response?
Seems simple enough, all the base models do is continuation, but you can do few-shot prompting to establish a pattern if you want. I will have to give it a try.