Set a framework for research. Return JSON {"question":"...","scope":"...","exclusions":"..."}. question—research question; scope — framework: period (for what years), place (country/region), depth (overview or in detail) and objects (what exactly are we considering); exclusions - which we clearly DO NOT consider in order to cut off the excess.
Mechanics: search without boundaries is like a net with holes - it pulls out everything, including junk and someone else's context. The model doesn't know what's "fresh" and "relevant" to you until you set the frame. The frame is the filter it applies to each source. Four frame axes. Time: “over the last 2 years” cuts off obsolete. Location: “in my country or city” removes irrelevant markets. Depth: “overview” versus “detailed with numbers” - this is a different amount of work. Objects: list what exactly we are comparing. Insider trick: write not only what to include, but also what to exclude. Explicit exclusions (“without corporate tariffs”, “without outdated models”) save half of the report and prevent the model from leaking to the side. The ban works more powerfully than another clarification. Second trick: if data quickly becomes outdated (prices, laws), always set the date of relevance - “for 2026”. Without it, the model will mix up the years and produce contradictions. A typical mistake: a broad request “everything about X” - you get a superficial mess from different years and countries.
Unlock access to submit solutions for instant AI review.