ResearchTopic

ResearchTopic

Single-pass parallel web research. Decompose, fan out, synthesize. The goal is a structured answer with citations and explicit confidence levels, returned in one cycle.

For multi-perspective research where specialists challenge each other across rounds, use ResearchCouncil.

Step 1: Decompose

Break the research question into 3-5 sub-queries covering different angles. Examples:

Original	Sub-queries
"best practices for tool X"	Official docs · GitHub issues + discussions · YouTube/Twitter creators using it · Reddit threads · Alternatives
"is library X production-ready"	Maintainer activity · Open critical issues · Production case studies · Recent security advisories
"landscape for AI code review tools"	Vendor offerings · Open-source TUI · Editor plugins · IDE-native features · Recent benchmarks

Decomposition is the highest-leverage step — a sloppy query becomes a confused report. Name the angles before spawning.

Step 2: Spawn parallel WebResearcher agents

For each sub-query, spawn a WebResearcher via the Task tool in a single message so they run in parallel. Each prompt includes the sub-question, brief context of the parent goal, an explicit ask for citations + confidence level, and a cap (~300 words per agent).

Step 3: Synthesize

Combine the agent reports into one structured output. Don't paste raw reports — synthesize:

## Research: <topic>

### Summary
One paragraph covering the load-bearing findings.

### Findings
1. **<finding>** — confidence: established | likely | uncertain. <One-sentence justification.> [source][1]
2. ...

### Conflicts
Where sources disagree, flag the disagreement and which side has more credibility.

### Gaps
What couldn't be determined; what would need primary-source access.

### What this means for you
(Only if the research was triggered by a decision question.) Concrete next steps.

### Sources
[1]: <url> "title"

Red Flags

Thought	Reality
"I'll just search directly without spawning agents"	Parallel fan-out is why this skill exists; in-context serial searches eat the context window.
"Top search results are good enough"	Vendor blogs, AI-summary spam, out-of-date posts rank well. Evaluate authority + recency.
"If only one source says X, drop it"	Often the most valuable signal. Flag as low-confidence, don't drop.
"Skip Conflicts if all sources agree"	Agreement is itself a finding — write "all sources concur on X" explicitly.