| name | Web Research |
| description | Use when you need current or external facts — searching the web, reading multiple sources, cross-checking claims, and synthesizing an answer with citations instead of guessing from memory. |
| version | 1.0.0 |
Web Research
Your training data is stale and incomplete. When a question turns on
current events, prices, versions, docs, or any fact you're not sure of,
look it up — and show where the answer came from.
When to reach for the web
- The answer could have changed since your knowledge cutoff — releases,
prices, news, who holds a role.
- The user asks about a specific product, library version, API, or company.
- You're about to state a number, date, or quote you can't verify from the
conversation.
- Do not search for things you already know reliably or that are already
in the repo or context.
Method
- Check for a search/fetch tool or a configured MCP server. If none is
available, say so and answer from what you know — flagged as unverified.
Do not invent a citation.
- Start broad, then narrow. Turn the question into 2–4 focused queries,
not one vague one.
- Open the primary source. Prefer official docs, the vendor's own site, or
the original report over a blog summarizing it.
- Cross-check anything surprising or load-bearing against a second,
independent source before you rely on it.
- Note the date on each source — a confident older page can be wrong now.
Reporting
- Cite each non-obvious claim with its source (title + link), and end with
a short "Sources" list.
- Separate what the sources say from your own inference. Don't launder a
guess as a finding.
- If sources conflict, say so and give the most credible reading rather
than averaging them.
See references/source-checklist.md.
Anti-patterns
- Answering a "what's the latest…" question from memory without checking.
- Citing a single source for a contested or high-stakes claim.
- Quoting a page you never actually opened.
- Dumping raw search results instead of synthesizing an answer.