web-research

Web Research

Ground answers in current, citable sources. The skill assumes the model already knows how to think — it provides ordering, source-quality heuristics, and reporting expectations only.

Workflow

1. Scope the question. If genuinely ambiguous, ask one targeted question. Skip clarification when the topic is self-contained (a specific library version, a named API, an identifiable product).
2. Search. Use WebSearch with precise terms. Run multiple queries in parallel when investigating distinct angles or comparing alternatives.
3. Fetch primary sources. Use WebFetch with a focused extraction prompt — pull only the slice you need, not the whole page. For GitHub URLs prefer gh (e.g. gh api, gh pr view) over WebFetch.
4. Cross-check claims that matter. Versions, security advisories, prices, breaking changes, current state of an API — one source is a lead, two corroborating sources is a fact.
5. Cite sources inline in the final answer so the user can verify.

Source quality

Prefer in this order:

• Official documentation, RFCs, standards bodies, vendor security advisories
• Project repositories, release notes, changelogs, issue trackers
• Reputable engineering blogs, conference talks, well-cited papers
• Stack Overflow / GitHub issues for symptom-driven debugging questions
• News outlets only when the question is genuinely news-driven

Discount AI-generated content farms, undated tutorials, SEO aggregator sites, and any source that cannot be tied to a person, org, or date.

Library docs: prefer Context7

For library/framework documentation (React, Django, FastStream, Vespa, anything imported in the codebase), the project ships Context7 MCP. Use mcp__plugin_context7_context7__resolve-library-id followed by query-docs instead of a generic web search — version-accurate docs return faster and without summarisation drift.

Fall back to WebSearch only when Context7 has no entry for the library or the question is about something Context7 does not cover (release notes, security advisories, community discussion, blog posts).

Reporting

Keep the answer focused on what the user asked. Default structure:

• Direct answer — one or two sentences resolving the question.
• Evidence — relevant findings, each tied to a source URL.
• Caveats — outdated info, conflicting sources, gaps in coverage.

Do not narrate the search process. The user wants the answer, not a transcript of how you found it.