[Docs] Prompt engineering strategy guide

[aaronvg]

1. Chain of thought
2. Add examples with structure you want
3. Reduce the input size
4. Identify contradicting statements in your prompts, like if you say “identify all facts and claims”. But then say “exclude opinions” later in the prompt, it should tell you.
5. Long inputs - use “here are all of the relevant facts”.
   1. “Json array with all the relevant facts\n print_type(output)\nAll facts or claims json:”
      • works if you’re trying to analyze every sentence
6. Enumerate every sentence you want the model to take a look at.
7. Symbol tuning
8. Use Delimiters for inputs and examples
9. Use “output json array” when the output is expected to be an array and “output json” when it is not.
10. Add a category that is an “out”.

Each example should be accompanied with runnable baml project and link to research papers