Parallel Research Skill

Parallel Research

Overview

When research scope is large, decompose into independent questions and spawn parallel researchers. This skill turns a single researcher into a research coordinator.

When to Use

When research question spans multiple distinct domains
When research would exceed context window if done sequentially
When independent aspects can be researched concurrently
When RPIV or Planner requests broad research coverage

Decomposition Criteria

A research question should be decomposed when:

| Criterion | Threshold | Example | |-----------|-----------|---------| | Topic breadth | 3+ distinct areas | "Research auth, caching, AND API patterns" | | Estimated context | >50% for single researcher | Large codebase analysis | | Independence | Aspects don't depend on each other | Security review vs performance review |

A research question should NOT be decomposed when:

Topics are deeply interrelated
Sequential discovery is required (A informs B)
Total scope fits in single context window

Decomposition Process

Step 1: Analyze Research Scope

Parse the research request to identify:

Core questions that must be answered
Independent aspects that can be parallelized
Dependencies between aspects (if any)

Step 2: Design Research Questions

For each independent aspect, create a focused research question:

## Research Question: [Aspect Name]

**Focus**: [Specific scope - what to research]
**Boundaries**: [What NOT to research - stay focused]
**Output**: [Expected deliverable format]
**Context needed**: [What context to provide]

Step 3: Spawn Parallel Researchers

Use the Task tool to spawn researchers:

task({
  subagent_type: "researcher",
  description: "Research [aspect name]",
  prompt: "Research Question: [question]
           Focus: [scope]
           Boundaries: [limits]
           Output format: [expected format]
           
           Return findings as structured Markdown with file:line references."
})

Spawn all independent researchers simultaneously by issuing multiple Task calls in parallel.

Step 4: Collect Results

As each researcher returns:

Validate the response addresses the question
Extract key findings
Note any cross-references or dependencies discovered

Step 5: Synthesize Findings

Combine all research into a unified output:

# Research Synthesis: [Original Topic]

## Executive Summary
[High-level findings across all research tracks]

## Detailed Findings

### [Aspect 1 Name]
[Synthesized findings from Researcher 1]
- Key discovery 1 (file:line)
- Key discovery 2 (file:line)

### [Aspect 2 Name]
[Synthesized findings from Researcher 2]
- Key discovery 1 (file:line)
- Key discovery 2 (file:line)

## Cross-Cutting Themes
[Patterns or connections discovered across research tracks]

## Recommendations
[Actionable next steps based on findings]

## References
[All file:line references from all researchers]

Output to Parent

When research is complete, return to orchestrator (RPIV or Planner):

RESEARCH_COMPLETE
Topics researched: N
Key findings:
  - [Finding 1]
  - [Finding 2]
Full synthesis: [path to research doc]

Error Handling

Researcher fails: Continue with other researchers, note gap in synthesis
Context overflow in child: Accept partial findings, note limitation
Conflicting findings: Document both perspectives, flag for human review

Agent Skills: Parallel Research

Install this agent skill to your local

Skill Files