MEPX
Chapter 8 of 10All chapters

Chapter 8 of 10

Evaluating prompts

Beyond it looked better.

Test sets

Collect ten to twenty representative inputs with notes on what a good answer contains. Run any prompt change against all of them before adopting it.

  • Single examples mislead because output varies between runs.
  • Record failures; they are the most informative cases.

Measuring

Score against criteria you wrote in advance. For production use, automated checks plus periodic human review is the practical arrangement.