test-code-guide

Test public behavior and contracts rather than internal implementation steps. Derive expected values independently of the implementation, mock only boundary collaborators like network and time, and verify the test still passes after behavior-preserving refactors.

Place each behavior at the lowest level capable of verifying it, since higher levels are slower and harder to debug. Follow the test pyramid ratio principle and reserve E2E tests for integration points and user journeys only.

Use equivalence partitioning with boundary value analysis for input ranges, pairwise combinatorial testing for parameter interactions, state transition testing for order-dependent logic, and decision tables for business rules. Pick techniques based on the defect hypothesis you are targeting.

Flakiness usually comes from hard-coded sleeps, real time or network dependencies, random values without fixed seeds, or shared state between tests. Replace sleeps with polling plus timeout or event subscription, inject clocks, and quarantine flaky tests for root-cause analysis instead of hiding them with retries.

Yes, the Skill is framework-agnostic and enforces no specific runner, assertion library, or directory convention. It first reads the target repository's existing tests to adopt its conventions, then applies framework-independent quality principles.

It does not trigger for requests that only run existing tests without writing or reviewing them. Full review-fix-commit iteration loops are delegated to a separate self-feedback-loop skill, and installing missing test infrastructure is handled as a separate request.