web-research

Web Research

Don't just dump search results. Synthesize.

Plan first

Before searching, write a one-line plan:

• What's the actual question? (Sometimes it's not what was literally asked.)
• What sources would be authoritative? (Official docs > vendor blog > community blog > random blog.)
• What would change the answer in the next 6 months? (If "yes," date-stamp the answer.)

Search

Use WebSearch for discovery, then WebFetch to read the most promising 2–4 sources directly. Don't read more than 5 — diminishing returns.

If the question is "compare A vs B," fetch the docs/landing pages for both and one independent comparison.

For npm/PyPI/crate freshness or maintenance status: use the package registry API or npm view <pkg> time, not a blog post.

Synthesize

Output template:

# {restated question}

**Short answer:** {one sentence, no hedging}

## Why
{2–4 sentences of the actual reasoning}

## Caveats
{anything that could flip the answer — version dependencies, recency, user's specific context}

## Sources
- [{title}]({url})
- ...

Calibration

• If you found <2 sources that agree, lower confidence in the short answer ("likely…", "as of {date}…").
• If sources disagree, list both positions and which one has the better evidence.
• If the question is unanswerable from the open web, say so plainly. Don't invent.

When to spawn a subagent instead

If the question genuinely needs >5 sources or branches into sub-questions → spawn the researcher subagent so the raw fetches don't pollute the main context.