- if you have good data (inputs and expected outputs), you can derive a prompt, optimised for the llm you're using, using that data
- if the llm changes, you just run the optimiser again
- you need a lot of good data to make it work, but you probably need that data for evals too
- changing dspy signatures over time is more intuitive than extending prompt texts and versioning them
- dspy uses a variety of prompt optimisation techniques under the hood which you get for free
  - REACT, COT, GEPA etc

The value proposition of dspy