feedback loop is my favorite! Also it helps to create an md file to add all the verification and failures which may happen otherwise. AI has a tendency to happy ship features and while looking at verification (which is created by its own) it keeps a check. For example: "Change the login button blue in color and also update the verification.md file with possible failures which may not get this job done or get the job done incorrectly. Review those verifications and then only finish the user request"