Research Command

Orchestrate comprehensive research using multiple parallel sub-agents.

Validation

If no topic provided, ask for one.

Research Flow

1. Analyze Request

Before proposing a research plan:

• Understand the user's question/topic
• Consider the project context (CLAUDE.md, tech stack, existing patterns)
• Review conversation history for relevant context
• Identify knowledge gaps that need filling

2. Propose Research Plan

Suggest a tailored research plan. Dynamically determine which research agents are needed:

Possible Research Dimensions:

Dimension	When to Include
Codebase - Existing Patterns	Feature touches existing code
Codebase - Architecture	Architectural decision needed
Codebase - Frontend	UI/UX or frontend changes
Codebase - Backend	API/database/server changes
Codebase - Infrastructure	DevOps, deployment, CI/CD
Documentation - Framework	Using framework features
Documentation - Libraries	Using external libraries
Documentation - APIs	Integrating external services
Community - Best Practices	Common patterns and anti-patterns
Community - Recent Changes	Library updates, deprecations

Output Format:

## Research Plan: [topic]

Based on your question and this project's context, I recommend:

### Proposed Research Agents (N)

1. **[Category] [Specific Focus]**
   - What: [What this agent will investigate]
   - Why: [Why this is relevant to your question]

2. **[Category] [Specific Focus]**
   - What: [description]
   - Why: [relevance]

[...more as needed...]

### Questions Before Proceeding

- [Any clarifying question if the request is ambiguous]

---
Proceed with this plan? (yes / adjust / add more / cancel)

3. Await Approval

Wait for user to:

• Approve ("yes", "proceed", "looks good")
• Adjust ("skip the community one", "also check X")
• Add ("also research Y")
• Cancel ("nevermind")

4. Execute Research

Once approved, spawn agents using the Task tool:

For EACH approved research dimension:
  - Use Task tool with appropriate subagent_type
  - Prompt should specify exactly what to find
  - Request structured findings in return

Subagent mapping:

• Codebase searches → subagent_type: "Explore"
• Documentation → subagent_type: "documentation-researcher"
• Community/best practices → subagent_type: "documentation-researcher" with web search

IMPORTANT: Spawn all agents in a SINGLE message with multiple Task tool calls for parallel execution.

5. Synthesize Findings

After all agents return:

Output Format:

## Research Findings: [topic]

### Summary
[2-3 sentence executive summary]

### Key Discoveries

#### From Codebase
- [Finding 1 with file:line references]
- [Finding 2]

#### From Documentation
- [Finding 1 with source links]
- [Finding 2]

#### From Community
- [Best practice or pattern]
- [Common pitfall to avoid]

### Recommendations
1. [Actionable recommendation]
2. [Another recommendation]

### Next Steps
- [ ] [Suggested action]
- [ ] [Another action]

---
Save these findings? (`/progress` to create knowledge document)

Example

User: /research "how should we implement caching"

Claude Response:

## Research Plan: Implementing Caching

Based on your Next.js + PostgreSQL project, I recommend:

### Proposed Research Agents (4)

1. **Codebase - Existing Patterns**
   - What: Search for any existing caching (Redis, in-memory, HTTP cache headers)
   - Why: Avoid duplicating or conflicting with existing caching

2. **Codebase - Data Access Layer**
   - What: Understand current database queries and ORM patterns
   - Why: Identify cache invalidation points

3. **Documentation - Next.js Caching**
   - What: Next.js 14 caching strategies (ISR, Route Handlers, fetch cache)
   - Why: Use framework-native caching before adding external solutions

4. **Community - Best Practices**
   - What: Redis vs in-memory vs CDN caching patterns for your stack
   - Why: Learn from others' mistakes and successes

---
Proceed with this plan?

Notes

• Prefer fewer, more focused agents over many broad ones
• 3-5 agents is typical; more than 7 is probably too broad
• Each agent should have a clear, specific mission
• Always synthesize findings - don't just dump raw results