The most expensive AI mistake isn’t a syntax error. It silently picks an interpretation where the task is ambiguous.
“Add notifications” is a dozen questions at once: who to send them to, when, what counts as a duplicate, what to do if delivery fails, whether user settings are needed.
So on ambiguous tasks I don’t ask for code in the first message. First I ask for an assumptions log: what the model considers known, what it assumes and which questions block the implementation.
Don’t write code. Break down the task: — facts from the description; — your assumptions; — questions without which you can’t start; — safe default values. Separate blocking questions from non-blocking ones.
After that it’s easier to see where AI has “made up” the business logic for you. And such a list is convenient to agree on with the product manager before plausible but wrong code appears in the repository.
The technique is especially useful for integrations, access rights and tasks with a vague spec. It isn’t needed if you’re changing a local function with clear inputs and tests: there an extra dialogue only slows you down.