Prompt Regression Testing

Core Workflow

1. Identify the behavior that must not regress and the artifact it affects.
2. Create compact prompts for normal, edge, and known-failure cases.
3. Define expected invariants rather than brittle exact wording.
4. Pair each prompt with acceptance criteria and expected routing.
5. Add negative cases for over-triggering, unsafe actions, or missing context.
6. Record when the prompt should be re-run: before release, after skill edits, after host behavior changes, or after incidents.

Safety Rules

• Do not include secrets, real customer data, or sensitive incidents in committed prompt tests.
• Do not make tests depend on exact prose unless exact language is the product.
• Do not rely only on happy-path prompts for high-risk skills.