I agree this is an interesting direction, I think this is on the roadmap for DSPy [https://github.com/stanfordnlp/dspy https://github.com/stanfordnlp/dspy], but

I agree this is an interesting direction, I think this is on the roadmap for DSPy [https://github.com/stanfordnlp/dspy https://github.com/stanfordnlp/dspy], but right now they mainly focus on optimizing the in-context examples.