Key answer
AI can help clarify requirements, but it need not question you until nothing remains uncertain. Prioritize information that changes the deliverable, audience or tradeoffs; decide other details after seeing a draft.
Full explanation
CLAM validated selective clarification: detect ambiguity, ask when necessary, and answer using the additional information. It studied factual QA with simulated users who knew the intended answers, mainly evaluating one clarification round. It supports appropriate questions resolving ambiguity, not endless questioning discovering every unformed goal. Original study
For requirement discovery, describe what to improve and where the result will be used; request two or three directions and their differences; select the closest and answer consequential questions; then confirm current requirements. This is beginner-oriented organization, not a fixed four-step procedure validated by that paper.
I want to improve my website's beginner experience but have not chosen what to build. Suggest possible goals and the problems each addresses, then ask the questions that best distinguish my needs. Do not treat suggestions as approved requirements. Once enough information exists for proposals, summarize confirmed points for me to choose.
“This example is close, but I do not want…” is also useful. Dislikes and actual sticking points narrow the scope.
This preceding conversation does not necessarily reduce later quality merely because it contains questions and clarification. Multi-turn research found some models struggled with requirements received piecemeal, sometimes mitigated by summaries. It does not prove every accurate clarification process harmful. Before execution, collect confirmed goals, retained conditions and undecided items; do not convert rejected guesses into final requirements. Section 7.1