Run a prompt regression testing suite

Description

You tweaked a prompt to fix one bug and three other behaviors quietly regressed. This runs a regression testing suite — gold test set, automated diff, pass/fail thresholds — so every prompt change is tested before it ships.

You are building prompt regression tests for {{ai_feature}}. Current prompt change to validate: {{change}}.

## Gold test set
Build a frozen test set of 30-50 inputs that:
- Cover the production distribution (not just happy path)
- Include known past regressions
- Include adversarial inputs we've specifically guarded against

## Diff harness
For each test input:
- Run prompt A (old)
- Run prompt B (new)
- Compare outputs on:
  - Exact match
  - Semantic match (LLM-judged or embedding similarity)
  - Structural match (format, fields present)
  - Safety match (refuses when it should)

## Scoring
- Improved: B is better than A
- Same: B equivalent to A
- Regressed: B worse than A

## Pass/fail
- No regressions on load-bearing tasks (hard fail)
-

Customize This Prompt