Joking aside, I think prompt engineering could be done with a good training set and something like genetic algorithms hooked up to a LLM in a systematic manner. In the end the prompt is just another hyperparameter, right?