Most failures with an AI happen before the first prompt: you ask for what seems needed, not for what is. The model answers the question asked, and a question that is not about the goal leads away from it.
Every question asserts more than it asks. “Which of the three frameworks?” asserts that a framework is needed, that there are three and that one fits; if any of that is false, every direct answer is wrong however good it sounds. Before asking, answer three things: what I actually want, how I will know I have it, and what I already take for granted. Write the presuppositions down and check first the one whose failure would change the decision most.
Same rule in code: an XY problem in a ticket is a presupposition nobody checked. Initiative — what counts as the problem — and the signature under the result stay with the human; this is a norm we choose, not a fact about the model.
Where it comes from. The engineering name is the XY problem: someone wants X, believes Y is the way to get it, and asks about Y. Eric S. Raymond’s “How To Ask Questions The Smart Way” gives the rule as “Describe the goal, not the step”. The idea of a hidden presupposition is older and comes from the philosophy of language — Frege and later Strawson on statements that presuppose something they do not assert — and the everyday form is the loaded question. Gause and Weinberg’s Are Your Lights On? (1982) is a short book-length treatment of “what is the problem, really”.
An engineering example. “How do I make the consumer skip a bad message faster?” presupposes that skipping is acceptable, that the message is bad rather than the parser, and that speed is the issue. The goal is usually “the partition keeps flowing and no record is silently lost”, which leads to a dead-letter queue with an owner (81) — an answer the original question could never have produced. Stated up front, the goal also gives the done criterion: lag back under the SLO, and every skipped record accounted for.
Where it is overkill. When the goal is known and narrow — a syntax question, a flag name — interrogating presuppositions is waste (111). And in open exploration the done criterion cannot be stated in advance; then state instead what result would change your direction, and revisit the goal when one arrives.