How do you evaluate if a prompt is good enough?

Answer preview

Build eval set: 50-100 real inputs with expected outputs (golden dataset).…