| name | testing-craft |
| description | Testing standards — what to test, at which seam, and what makes a test worth having. Use whenever writing or reviewing tests for new behavior. |
| user-invocable | false |
Testing Craft
What a good test is
A good test reads like a specification of behavior and survives refactoring:
it exercises a public seam (API endpoint, exported function/hook, component
behavior, CLI) and asserts observable outcomes — not internals.
Where to test (choose the cheapest seam that gives a real signal)
- Unit: pure logic — parsers, calculations, reducers, validators.
- Integration: endpoint/service with real (or containerized/in-memory) DB — the
default for backend behavior.
- Component: render + interact + assert visible result (Testing Library style) —
the default for frontend behavior.
- E2E: only for the few critical user journeys; keep the suite small and stable.
Minimum coverage for new behavior
- Happy path with realistic data.
- One boundary/edge case that matters (empty, max, unicode, zero).
- Failure path: invalid input rejected with the right error; authz denied for
another user's resource (backend); error state rendered (frontend).
Anti-patterns (a reviewer will flag these as findings)
- Tautological assertions — recomputing the expected value with the same code path under test.
- Mocking the thing under test or mocking internals so deeply the test asserts
the mock, not the behavior.
- Tests coupled to implementation details (private method calls, CSS class names,
exact log strings) that break on any refactor.
- Asserting nothing (
expect(result).toBeDefined() as the only assertion).
- Sleeping/arbitrary timeouts instead of awaiting a deterministic signal — flaky by design.
- Deleting or skipping a failing test to make the suite green.
Discipline
- Run the whole relevant suite before declaring done, not only your new tests.
- Regression tests for bug fixes are written to fail on the bug first, then pass with the fix.
- Follow the project's existing test framework, helpers, and naming — don't introduce a second stack.
