Back in December, the handit team and us tried this methodology. They ingested around 40k calls to a given prompt and then handit A/B tested the prompt with an improved version. We tracked the performance and it went from 40% to 70%. Amazing work!!! keep it up!
