research

Research

Search the web for documentation and technical information relevant to a task using parallel agents.

Input

TASK: $1 DOCUMENTATION_URLS: $2

Workflow

1. Identify Search Angles

Based on the task, determine 2-3 distinct search angles:

• Official docs: Framework/library documentation, API references
• Community: Stack Overflow, GitHub issues, blog posts with solutions
• Related: Similar implementations, migration guides, changelog entries

2. Dispatch Parallel Research Agents

Launch 2 agents by default. Add a third if DOCUMENTATION_URLS were provided or if more than 2 clearly distinct search angles exist. Dispatch all in a single message (parallel execution). Use model: "sonnet" for each agent.

Each agent prompt MUST include:

• The full task description
• "You are ONLY searching for relevant documentation. Do NOT implement anything."
• "Use WebSearch to find sources, then WebFetch to read the most promising results"
• If DOCUMENTATION_URLS were provided, one agent should prioritize fetching and analyzing those specific URLs
• "Return structured findings in this format:"

## <Topic/Source>
**URLs**: <url1>, <url2>
**Summary**: <brief summary of what was found and why it's relevant>
**Key Details**: <specific code patterns, API signatures, config values, etc.>

3. Collect Results

• Merge findings from all agents
• Group related URLs together (e.g., multiple pages from same docs site)
• Drop irrelevant or low-quality results
• Prioritize: official docs > GitHub issues/PRs > blog posts > forum answers

4. Save Report

Create .claude/doc_search/ directory if it doesn't exist. Save to .claude/doc_search/<brief-task-slug>-research.md:

# Research Report: <brief task summary>

## Task
<task description>

## Findings

### <Topic 1>
**Sources**: <urls>
**Summary**: ...
**Key Details**: ...

### <Topic 2>
...

5. Return

Report the saved file path and a brief summary of key findings.