One task at a time, tests that check behaviour rather than implementation, and a review where AI reports but never fixes.