Take your everyday task (purchase, choice, analysis of a topic). Return JSON {"simple_question":"...","research_task":"...","why_deep":"..."}. simple_question - as a beginner would ask in one line; research_task - the same topic, rewritten as a research problem with a goal and at least one frame; why_deep — 1-2 sentences why deep research is useful here, and not a quick answer.
Mechanics: a common question for a neural network is one pass. The model answers from what it “remembers” and is often confidently wrong. The deep research mode works differently: it breaks the task into sub-questions, does a lot of searching across different sources, verifies what it finds and collects a structured report with conclusions and links. An analogy: a simple question is how to ask a friend “what to buy?”, and deep research is how to hire an assistant who will call stores, compare prices and bring a table. Technique: to activate the research mode, formulate not a question, but a task - with a goal and a framework. "Which is better?" gives chatter; “Compare 3 options based on price, reliability and timing and draw a conclusion for a beginner,” the report says. Insider: the narrower and more specific the frame, the deeper the model digs—it smears a broad request superficially. A typical mistake: ask in one line and wait for a miracle. Without purpose, criteria and format, you will get a general retelling, not research.