Use this skill when choosing what tests to add, update, or trust for a feature, fix, refactor, or AI-generated code change.
References
- Source synthesis: BettaTech, "¿Estamos haciendo MAL los tests de software?".
- Complementary strict review lens: thermo-nuclear-code-quality-review, especially for spotting test suites that preserve incidental complexity, over-mock brittle internals, or bless spaghetti growth.
Native Integration
Use the active agent's native terminal, test runner, browser, debugger, coverage, container, and CI tooling for the mechanics. This skill owns the testing decisions: what uncertainty matters, which behavioral contract should be locked down, and which test level gives the strongest useful signal with acceptable maintenance cost.
Principles
- Tests are a safety net for future change, not just a present-tense happy-path check.
- Prefer public behavior and stable contracts over implementation details.
- Manual testing does not scale as the system grows; automate repeatable confidence checks.
- Use unit tests for isolated pure logic, integration tests for component or service contracts, and end-to-end tests for critical user journeys.
- Prefer integration tests against real empty dependencies, containers, or local services when mocks would hide contract drift.
- Use mocks sparingly and deliberately; every mock is a maintenance promise about another system's behavior.
- Treat coverage as a diagnostic signal, not a target. High coverage with weak assertions is false confidence.
- Prioritize failure paths, edge cases, boundary conditions, and regressions before adding redundant happy-path coverage.
- Treat tests as the executable contract that constrains AI-generated changes.
Workflow
-
Name the uncertainty.
- State the behavior, user path, API contract, data invariant, or regression that could break.
- Identify who or what relies on that contract.
-
Choose the highest useful test level.
- Prefer an integration or endpoint-level test when it can verify the real contract without excessive setup.
- Use a unit test when the behavior is pure, isolated, or hard to reach through a higher-level seam.
- Use an end-to-end test when the risk lives in real user workflow wiring, browser behavior, auth, routing, or cross-service coordination.
-
Decide what must stay real.
- Keep databases, queues, filesystems, HTTP services, or time real when the contract with that dependency is the risk.
- Use containers, test databases, fixtures, or local fakes before broad mocks when practical.
- If a mock is necessary, document the reason in the test name, fixture, or nearby setup.
-
Cover the risk matrix.
- Start with the most important error path, boundary, permission failure, invalid input, race, or regression case.
- Add the happy path after the behavior's failure modes are constrained.
- Avoid adding tests that only execute lines without asserting meaningful outcomes.
-
Run the smallest meaningful loop.
- Run the focused test and confirm it fails for the expected reason when adding coverage.
- Make the smallest change needed to pass.
- Keep the focused check green while cleaning up names, fixtures, and setup.
-
Review test quality.
- Check that assertions describe outcomes, not internal steps.
- Remove brittle coupling to private helpers, incidental file layout, or arbitrary call counts.
- Watch for structural smells from the thermo-nuclear review lens: sprawling setup, special-case branches, cast-heavy fixtures, thin wrappers, or tests that force poor production abstractions.
Done When
- The test suite reduces a named uncertainty that matters to users, callers, or maintainers.
- The chosen test level matches the risk and avoids unnecessary brittleness.
- External contracts are verified with real dependencies where feasible, or mocks are explicit and justified.
- Error paths, edge cases, and regressions are covered before shallow coverage targets.
- The final checks pass and the remaining testing gaps are stated plainly.