maxqian888/cognia-next/web-research

The value of web research is bringing back facts the user can trust. A confident answer assembled from search snippets you never opened is worse than no answer — it looks authoritative and can be wrong.

Use the host-projected web_search and web_fetch tools. If search is unavailable, work only from URLs the user supplied or explain the limitation; never route a query through an undeclared network tool. Keep the report in the user's language.

Query with intent

• Make queries specific: include the version, the year, the exact error string, the proper noun. Broad queries return broad noise.
• Run a couple of differently-phrased searches rather than one — the best source often isn't on the first results page for your first phrasing.
• Know what you're looking for before you search. "What does X do" and "why is X broken" need different queries and different sources.

Read the source, not just the snippet

• Open the pages that look authoritative; don't answer from the search-result summary alone. Snippets are truncated, dated, and sometimes contradict the page body.
• Fetch with an extraction goal in mind — pull the specific fact or section you need, rather than asking for a vague summary of a long page.
• Prefer primary and official sources (official docs, the project's own repo, standards) over aggregators and tutorials. Note publication dates; a 2019 answer about a fast-moving tool may be stale.

Corroborate and attribute

• For a claim that matters, confirm it appears in more than one independent source. A single blog post is a lead, not a fact.
• Cite where each load-bearing claim came from — name the source or link it — so the user can verify. Distinguish what the sources state from what you're inferring by connecting them.
• When sources disagree or you couldn't confirm something, say so. "I found conflicting answers" or "I couldn't verify this" is honest and useful; a smoothed-over guess is not.

End substantive research with the sources you relied on, so the trail is visible.

For the source trust hierarchy, recency checks, and the corroboration rubric, see references/source-evaluation.md.