| name | web-research |
| description | Research a question on the open web and return a verified, cited answer. Trigger when the user asks to "look this up", "research X", "find out what/whether", "what's the latest on", or any question you can't answer reliably from memory — versions, prices, current events, docs, comparisons. Emphasizes multiple independent sources, checking dates, distinguishing primary from hearsay, and citing what you claim. |
web-research
The goal is a defensible answer, not the first plausible link. A single
source is a claim; two independent sources agreeing is evidence. Your job is to
turn a question into a small set of trustworthy findings, each traceable to
where it came from.
Use the web_search tool to find sources and web_fetch to read them in full —
snippets lie by omission; open the page before you quote it.
Loop
- Sharpen the question. What exact fact settles it? "Is library X
compatible with Y" → "does X's docs/changelog list Y support, and as of which
version". A vague question yields vague searches.
- Search deliberately. Start broad to map the terrain, then narrow with
specific terms, error strings, version numbers, or
site: filters. Rephrase
when results are thin — different words surface different sources. Prefer the
primary source: official docs, the changelog, the standard, the paper, the
vendor's own page — over a blog summarizing it.
- Open and read.
web_fetch the promising results. Read enough to confirm
the claim in context, not just a matching sentence.
- Cross-check. Confirm anything load-bearing against a second independent
source. Two blogs both citing the same original are one source. Watch for
an answer that everyone copied from one wrong post.
- Synthesize with citations. Answer the question directly, then support it
with links. Note your confidence and any disagreement you found.
Judging a source
- Primary > secondary > hearsay. Docs and specs beat a tutorial; a tutorial
beats a forum guess.
- Check the date. Software, prices, and events go stale fast. An accurate
2021 answer can be wrong today. Prefer dated pages; state the date you relied
on. Beware undated posts and content-farm SEO pages.
- Who benefits. A vendor comparing itself to a rival, an affiliate "best of"
list, marketing dressed as a benchmark — read for bias, corroborate elsewhere.
- Watch versions. "How to do X" often changes between major versions. Match
the answer to the version the user is on.
Reporting
- Lead with the answer. One or two sentences that actually resolve the
question, then the support.
- Cite inline. Attach a URL to each non-obvious claim so the user can
verify. Prefer a deep link to the exact page over a homepage.
- State confidence and gaps. "Confirmed by docs + changelog" vs. "one
undated blog, unverified". If sources conflict, say so and show both — don't
silently average them into a false consensus.
- Don't invent citations. Never attach a URL you didn't open, and never
paraphrase a source into saying more than it does. If the web didn't answer
it, say the web didn't answer it.
Don't
- Don't answer version/price/current-event questions from memory — search.
- Don't stop at the search snippet; the full page often contradicts it.
- Don't treat volume as truth — ten copies of one rumor is still one rumor.