← Prompt Engineering That Pays
Lesson 5 — Testing and Iterating Like a Scientist
One good output proves nothing. A fixed test set and a score tell you what actually improved.
♪ Download audio
Prefer to listen? Take this lesson on the commute.
- Never judge a prompt change on a single output; models vary and you are a biased grader.
- Freeze a test set of ten real inputs, including two hard cases, and reuse it for every version.
- Change exactly one variable per version and log it beside the score, or you learn nothing.
- Inspect per-item results, not just the average, so a rising mean does not hide a regression.
Loading the quiz…