| name | web-research |
| description | A method for multi-step web research that produces a cross-checked, cited answer instead of a single-search guess. Use for questions needing current, authoritative, or contested information. |
| metadata | {"version":"1.0"} |
Web research
Use this when a question needs current, authoritative, or contested information — not for things already known or in local context. The goal is a synthesized answer backed by independent, cited sources, not the first search result.
Method
-
Frame the question. Write down exactly what's being asked and what a complete answer must include. If it's broad or ambiguous, decompose it into sub-questions and resolve each — don't research a vague blob.
-
Plan sources before searching. For each sub-question, note what kind of source would actually settle it (primary docs, official changelog, standards body, peer-reviewed work, vendor docs, reputable reporting). Prefer primary sources over aggregators.
-
Search iteratively. Start broad to map the landscape, then narrow with specific terms you learned from the first pass. Reformulate when results are thin — try the precise product/version name, error string, or domain term rather than paraphrase.
-
Read, don't skim-and-assume. Open the actual source. Distinguish the claim from the source's spin. Note publication dates — stale info is a common trap, especially for fast-moving tools.
-
Cross-check every load-bearing claim against at least one independent source. Two pages copying the same press release are one source. If sources conflict, surface the conflict and weigh credibility/recency rather than picking silently.
-
Track provenance as you go. Keep each fact tied to its URL so the final answer can cite. An uncited claim is a guess.
-
Synthesize. Answer the original question directly first, then support it. Separate what's well-established from what's uncertain or contested. State what you could not confirm.
Output
- Lead with the direct answer / recommendation.
- Support each material claim with a citation (plain URL).
- Call out uncertainty, conflicts, and recency explicitly — don't smooth them over.
- If the question couldn't be fully answered, say what's missing and what would resolve it.
Pitfalls
- Stopping at the first plausible result.
- Treating volume of agreement as truth when it's all one origin.
- Ignoring dates on time-sensitive topics.
- Presenting a confident synthesis that hides a real source conflict.
- Paraphrasing a source's claim as fact without verifying it.