| name | research-sweep |
| description | Run a multi-source research sweep on a topic and synthesize a cited summary. ALWAYS invoke when the user says "research X", "sweep X", "what's being said about X", or asks for a grounded summary across sources before drafting something. NOT for a single-source lookup (just fetch that source).
|
| model | opus |
| allowed-tools | Read, WebSearch, WebFetch, Write |
| disable-model-invocation | false |
research-sweep
The spine of the procedure lives here; the per-source query patterns and the
synthesis spec live in references/, loaded only when the task reaches that
branch. That is progressive disclosure inside a skill — this body stays short.
Procedure
- Scope. If the topic is underspecified (no audience, timeframe, or angle),
ask one or two clarifying questions before sweeping. A vague sweep wastes the
whole run.
- Gather. Run the per-source queries. The patterns for each source (search,
forums, primary docs) are in
references/sources.md.
Read that file now.
- Verify. Cross-check any claim that appears in only one source before it
goes in the summary. A single-source claim is flagged, not asserted.
- Synthesize. Follow the two-pass spec in
references/synthesis.md — a mechanical rollup pass,
then a judgment pass. Read that file when you reach this step.
- Write the summary with inline citations to the URLs actually used, and a
short "what's still uncertain" section. Save it where the user asked, or
surface it inline if they did not specify.
Notes
- Cross-platform signals (a claim that shows up in several independent sources)
are weighted highest. A single loud source is the weakest evidence.
- Never fabricate engagement numbers or citation counts. If a number isn't in the
fetched content, don't put it in the summary.
