Research Report

Research reports

Research quality comes from separating what sources say from what you conclude, and making that separation visible.

Method

1. Frame the question in one sentence with its scope ("Which three competitors ship agentic spreadsheet features as of this month?"). Split into 3–6 sub-questions; if the user's ask is vague, state your reading of it before searching.
2. Gather. web_search per sub-question, web_fetch only the pages that plausibly answer it. For each source capture: claim, date, and how authoritative it is (primary doc > vendor post > press > blog).
3. Cross-check. A finding shared by two independent source classes may be stated plainly; anything single-sourced gets "reported" framing. Conflicts between sources are themselves findings — surface them.
4. Synthesize into structure: findings ordered by decision-relevance, not discovery order. Grade each: Established (multiple strong sources), Reported (single source), Inference (your reasoning — label it).
5. Deliver: long-form via a python-docx build script (document-report anatomy); comparative data via openpyxl with a Sources sheet; quick answers stay in chat. Call deliver on each file. Include a Sources section listing title, publisher, date, URL.

Honesty rules

• Never let a searched fact appear without a source in the deliverable.
• Dates matter: say when each claim was true ("as of the source's date").
• If searches can't answer the question, say exactly what was tried and what came back thin — a precise gap beats confident filler.

Pacing

For multi-hour investigations: create a goal, work sub-question by sub-question, and report interim findings in chat before assembling the final document.