prompt-regression-tests

Steps

1. Snapshot current outputs on the golden set.
2. Diff new outputs against snapshots on structure, not wording.
3. Flag semantic drift with a judge comparing old vs new.
4. Record results in the registry test format.

Done When

• [ ] Test record is dated and linked from the entry

Registry

Links

• Up: AI Engineer skills
• Siblings: llm-eval-harness, model-selection-eval
• Frameworks: react, few-shot, risen, xml-structured, chain-of-thought
• Tests: testing guide