research

Prompt file imported from mmarko88/mmarko88.github.io (.windsurf/workflows/research.md). Copyright stays with the author.

Research a question online

Usage: /research <question or topic>.

Implements the research-agent practices in AGENT_RESEARCH.md (§3 + the deep-research skill sources). Output is a conclusion with evidence, not a pile of links.

Steps

1. Sharpen the question. Before searching, pin down: the decision behind the question, the constraints that disqualify options, and what form the answer takes (recommendation, comparison, number, yes/no). If the sharpened question differs materially from what was asked, state the difference and confirm before spending effort. Pick depth: quick (1 authoritative source) / standard (2–3 cross-referenced) / deep (multi-hop, contradictions resolved).

2. Plan. todo_list the sub-questions — one focused aspect each. Simple fact-finding: 1 todo, no decomposition. Multi-faceted: one todo per facet so searches can run in parallel.

3. Search in bursts, from several angles. One phrasing returns one neighbourhood of results — vary:

   • vocabulary — practitioners, vendors, and academics name the same thing differently
   • stance — search the failure mode too ("X problems", "migrating away from X", "X postmortem"); critical sources are systematically underrepresented because vendors publish more than victims
   • recency — what was true three years ago may not be

   Two-stage filter: judge by title/snippet first, read_url_content only the promising ones. Prefer domain filters on search_web for primary sources (vendor docs, arxiv, official papers).

4. Weigh sources, don't count them. Ten blogs repeating one benchmark = one piece of evidence. Trace claims to their origin. Rank by: primary over secondary (docs, source, original paper); interest (who benefits if you believe this); specificity (states versions/conditions/hardware); falsifiability. SEO filler, listicles, and anonymous claims are leads, not evidence.

5. Stop conditions — stop searching when: the claim set has required coverage AND marginal value drops (new results are duplicates), OR hard cap ~3 search rounds. Detect loops: if queries start repeating or rephrasing yields the same domains, pivot stance or stop — don't loop.

6. Resolve disagreement, don't average it. Conflicts are the most informative moment. Usually: different conditions (both right at different scales/versions — find the boundary), different definitions, one is stale, or one is wrong (say so, with reason). Record unresolved conflicts; never force consensus.

7. Report conclusion-first. Quick questions: answer inline. Standard/deep: write RESEARCH-<topic-slug>.md at repo root:

   # <Question>
   **Answer:** conclusion in 2–3 sentences.
   **Confidence:** high / medium / low — and what drives it.
   
   ## Why — reasoning, evidence attached to each step [1]
   ## What would change this — the finding that would flip the conclusion
   ## Considered and rejected — alternatives + why each lost
   ## Open — what's still unknown and what would resolve it
   
   ## Sources
   [1] Title: URL (one number per unique URL, sequential, no gaps)

   Rules: no self-referential language ("I found…"); every factual claim cited; never cite a source you didn't open; claims with <2 sources go in Open questions, not Findings; no "probably/likely" filling gaps.

8. Verify. Re-read the original request; confirm every aspect is addressed and the report stands alone. Present the conclusion to the user first, file path second.