Research Skill

Delegates research to general-purpose subagents to preserve the main context window. Supports parallel execution for multiple queries.

Workflow

1. Identify Research Queries

Break the user's request into discrete, focused research questions:

• GitHub repo questions → route to deepwiki
• Code/API/library questions → route to exa code context
• General web/current events → route to exa web search

2. Launch Parallel Subagents

CRITICAL: Launch multiple Task calls in a SINGLE message block for parallel execution.

Task(subagent_type="general-purpose", prompt="...")
Task(subagent_type="general-purpose", prompt="...")
Task(subagent_type="general-purpose", prompt="...")

Each subagent prompt MUST include:

1. The specific research question
2. Tool priority instructions (see below)
3. Request for concise summary with source URLs

3. Synthesize Results

Combine subagent outputs into a coherent response for the user.

4. Cite Sources

Include all URLs from research in a Sources section.

Subagent Prompt Template

Use this template for each research subagent:

Research: [SPECIFIC QUESTION]

Tool priority:
1. Exa (primary): Use `mcp__exa__web_search_exa` for web, `mcp__exa__get_code_context_exa` for code/APIs
2. Deepwiki (for GitHub repos): Use `mcp__deepwiki__ask_question` with format `owner/repo`
3. Skip using WebSearch

Return:
- Concise summary (2-4 paragraphs max)
- Key findings as bullet points
- Source URLs

Examples

Single Query

Task(
  subagent_type="general-purpose",
  prompt="Research: How does React Server Components handle data fetching?\n\nTool priority:\n1. Use mcp__exa__get_code_context_exa for code patterns\n2. Use mcp__deepwiki__ask_question with 'facebook/react' for official docs."
)

Parallel Queries (single message block)

Task(subagent_type="general-purpose", prompt="Research Next.js app router patterns using exa code context...")
Task(subagent_type="general-purpose", prompt="Research Vercel deployment config using deepwiki vercel/next.js...")
Task(subagent_type="general-purpose", prompt="Research edge runtime limitations using exa web search...")

Guidelines

• Prefer specific queries over broad ones
• Include language/framework context in prompts
• For deepwiki, always use owner/repo format
• Maximum 4-5 parallel subagents to avoid overwhelming
• Each subagent should return focused, actionable findings