Engineering notes / AI engineering practice
Keep AI development fast and reviewable
How clear specifications, focused changes and meaningful tests support responsible use of coding agents.
Start with an observable outcome
Before implementation, describe what the change should do, what it must preserve and how it will be checked. Give the coding agent the relevant interfaces and repository conventions. A bounded task is easier to review than an open-ended instruction to improve a system.
Use specifications and tests together
Specification-driven development establishes intent and constraints. Test-driven development turns expected behaviour into executable feedback where that approach is useful. Neither removes the need for judgement: tests can encode an incorrect assumption, and generated code can satisfy a narrow example while missing an integration problem.
Keep ownership with the engineer
Review changes in small increments. Check dependencies, data handling, failure paths and the fit with the existing architecture. Run the checks that establish the acceptance criteria, then inspect the actual user flow. AI can shorten the implementation loop; the engineer remains responsible for the decision to release.