Protect meaningful behavior with the smallest useful test set. Match the request: ordinary testing, regression work, or a requested test-first loop. A tests-only request keeps production changes outside its scope.
Contract
Read the relevant public surface, callers, existing tests and repository conventions. Identify the caller-visible contract, important failure risk and cheapest boundary that exercises it. Reuse accepted decisions; resolve only real contract ambiguity. Keep production APIs product-driven.
Minimum
Reuse checks that already protect the change. Add tests for distinct uncovered risks, not edited lines. Behavior-preserving refactors may need only existing checks. Prefer one readable test for one outcome, with the relevant assertions; split when independent behaviors or setups would obscure failures. Minimize redundancy without deleting distinct protection.
Mode
- Tests: exercise the subject with concrete inputs and independent expected outcomes. Choose the smallest realistic setup and true external boundary.
- TDD / regression: read Test-first for the red → green loop, failing-before evidence and after-fix sensitivity checks.
- Quality: also read Test quality before writing async tests or boundary doubles, and when pruning tests or evaluating suspect assertions, constants or overlapping coverage.
Proof
A useful test distinguishes a relevant defect while accepting equivalent implementations. Observe returned values, public state, persistence, events or boundary interactions when those interactions are the contract. Use real mechanisms or small faithful fakes where their behavior matters.
Tautological tests considered harmful. A result compared with itself has no independent oracle; obtain expectations from the contract or a worked example. Copying an algorithm can share its bug but is not necessarily a strictly tautological assertion.
Internal constant, prompt-string or source-shape pins usually protect no behavior; exercise the mechanism instead. Independently specified public values and absence of forbidden effects can be real contracts.
Evidence
Run focused checks and the relevant suite. Report actual commands, results and coverage gaps. Distinguish reproduction of the original bug, sensitivity to a mutation, and passing-after evidence. Missing required automated tests or execution remain pending; an alternative check does not waive the request.