OpenAI has some advice about it

Types

Can do SGD on the prompt embeddings.
Can use RL with discrete prompt sets
In The Wisdom of Hindsight Makes Language Models Better Instruction Followers, they say that “prompt engineering and finetuning have orthogonal benefits”

Tricks

Please criticize your previous answer and tell me ways in which it was unhelpful, or ways in which it could have been approached better.

Given this criticism, can you try answering again?

Promptsource has a collection of prompts for many different eval tasks:

‣ recommends giving it an explicit structure:

Here’s the prompt he uses to generate the plan: