Use this advisory skill for production product or business code. Prompt, skill, and agent-behavior evaluation need different methods.
Method
- Read the exported surface, its callers, nearby tests, and shared test utilities. State one caller-visible contract and the regression risk it guards.
- Write one test whose name states the expected outcome. Keep the decisive input visible at the call site; split distinct behaviors.
- Assert observable effects: returned values, public state, stable error types or codes, persistence, emitted events, rendered UI, or a boundary call when that call is itself the contract.
- Build the smallest realistic setup. Match real data and error shapes; isolate only the lowest true external boundary.
- For a regression or test-first change, confirm the test is red for the intended reason before relying on it.
- Run the focused test and relevant suite. Finish only when breaking the stated contract would fail the test while harmless internal refactoring would not.
Judgment
- Prefer exact domain outcomes and meaningful fields. Use full snapshots or full error messages only when the entire representation is the contract.
- Use a real dependency or small fake when its behavior matters. Mocks fit slow, flaky, costly, or external boundaries; verify collaborator calls only when the call is externally meaningful.
- Wait for the condition or event that proves async completion. Use a controlled clock or fixed delay only when timing itself is the contract.
- Keep contract-relevant fields in test doubles; partial mock shapes can hide downstream assumptions.
- Let builders remove noise while exposing the value that makes the case meaningful.
- Treat private-state access as a coupling signal. Keep test-only affordances in test utilities and production APIs product-driven.
- Match the repository's existing framework, naming, placement, and helper conventions.