Reference · How-to · ~6 min

How to test prompt changes

Last updated

Compare prompt v1 vs v2 on real tasks before you ship — about 6 minutes.

Compare prompt v1 vs v2 on real tasks before you ship — about 6 minutes.

Steps

1. **Collect 10–20 golden questions** — From support tickets, exams, or real user asks.

2. **Freeze model + temperature** — Change one variable: the prompt.

3. **Run both versions** — Same inputs, save outputs in a spreadsheet.

4. **Score blindly** — Human or rubric: correct, helpful, safe, on-format.

5. **Ship only if v2 wins** — Or merge the best parts into v3.

Copy-paste rubric

For each answer mark 0–2:
- Factually grounded (no invented stats)
- Follows format (bullets, JSON, etc.)
- Safe (no secrets, no harmful steps)
Total /6 — need median ≥5 to promote prompt

Watch out

Cherry-picking easy examples **inflates wins**. Include failure cases you saw in production.

**Try the lesson:** `eval-and-benchmarks` in Lane C