Almost every "the AI built the wrong thing" story I hear traces back to the same root cause, which is a decision that existed in someone's head and never made it into writing, so the model filled that gap with a reasonable guess and the team spent the night discovering which guesses were wrong. Writing a short specification before coding, covering your hard constraints, your testable requirements, your schema and endpoints, and your task order, removes most of that rework at the source because every future session starts from decisions that are already made and permanently recorded. I put together a longer breakdown of how I structure these four files and why each one matters, and I'm happy to drop it below if anyone finds it useful. (Image by myspec.dev/)