research（研究）

Research Skill

API Keys Required

This skill works best with these optional API keys configured in ~/.env:

Works without API keys:

• Claude-based research (uses built-in WebSearch)
• Basic web fetching (uses built-in WebFetch)

Workflow Routing

Multi-Source Research Workflows

When user requests comprehensive parallel research: Examples: "do research on X", "research this topic", "find information about Y", "investigate this subject" → READ: ${PAI_DIR}/skills/research/workflows/conduct.md → EXECUTE: Parallel multi-agent research using available researcher agents

When user requests Claude-based research (FREE - no API keys): Examples: "use claude for research", "claude research on X", "use websearch to research Y" → READ: ${PAI_DIR}/skills/research/workflows/claude-research.ts → EXECUTE: Intelligent query decomposition with Claude's WebSearch

When Claude WebSearch fails or returns poor results (fallback): Examples: WebSearch returned nothing, stale results, rate-limited, user says "use gemini" → EXECUTE: Launch gemini-researcher agent via Task tool with the failed query → NOTE: Also use as parallel second opinion when confidence is low

When user requests Perplexity research (requires PERPLEXITY_API_KEY): Examples: "use perplexity to research X", "perplexity research on Y" → READ: ${PAI_DIR}/skills/research/workflows/perplexity-research.ts → EXECUTE: Fast web search with query decomposition via Perplexity API

When user requests interview preparation: Examples: "prepare interview questions for X", "interview research on Y" → READ: ${PAI_DIR}/skills/research/workflows/interview-research.md → EXECUTE: Interview prep with diverse question generation

Content Retrieval Workflows

When user indicates difficulty accessing content: Examples: "can't get this content", "site is blocking me", "CAPTCHA blocking" → READ: ${PAI_DIR}/skills/research/workflows/retrieve.md → EXECUTE: Content retrieval via WebFetch

When user provides YouTube URL: Examples: "get this youtube video", "extract from youtube URL" → READ: ${PAI_DIR}/skills/research/workflows/youtube-extraction.md → EXECUTE: YouTube content extraction

When user requests web scraping: Examples: "scrape this site", "extract data from this website" → READ: ${PAI_DIR}/skills/research/workflows/web-scraping.md → EXECUTE: Web scraping techniques and tools

Content Enhancement Workflows

When user requests content enhancement: Examples: "enhance this content", "improve this draft" → READ: ${PAI_DIR}/skills/research/workflows/enhance.md → EXECUTE: Content improvement and refinement

When user requests knowledge extraction: Examples: "extract knowledge from X", "get insights from this" → READ: ${PAI_DIR}/skills/research/workflows/extract-knowledge.md → EXECUTE: Knowledge extraction and synthesis

Multi-Source Research

Three Research Modes

QUICK RESEARCH MODE:

• User says "quick research" → Launch 1 agent per researcher type
• Timeout: 2 minutes
• Best for: Simple queries, straightforward questions

STANDARD RESEARCH MODE (Default):

• Default for most research requests → Launch 3 agents per researcher type
• Timeout: 3 minutes
• Best for: Most research needs, comprehensive coverage

EXTENSIVE RESEARCH MODE:

• User says "extensive research" → Launch 8 agents per researcher type
• Timeout: 10 minutes
• Best for: Deep-dive research, comprehensive reports

Available Research Agents

Check ${PAI_DIR}/agents/ for agents with "researcher" in their name:

• gemini-researcher - Uses Gemini CLI as fallback when WebSearch fails (requires GOOGLE_API_KEY)

Speed Benefits

• ❌ Old approach: Sequential searches → 5-10 minutes
• ✅ Quick mode: 1 agent per type → 2 minute timeout
• ✅ Standard mode: 3 agents per type → 3 minute timeout
• ✅ Extensive mode: 8 agents per type → 10 minute timeout

Intelligent Content Retrieval

Three-Layer Escalation System

Layer 1: Built-in Tools (Try First - FREE)

• WebFetch - Standard web content fetching
• WebSearch - Search engine queries
• When to use: Default for all content retrieval

Critical Rules:

• Always try simplest approach first (Layer 1)
• Escalate only when previous layer fails
• Document which layers were used and why

File Organization

Working Directory (Scratchpad)

${PAI_DIR}/scratchpad/YYYY-MM-DD-HHMMSS_research-[topic]/
├── raw-outputs/
├── synthesis-notes.md
└── draft-report.md

Permanent Storage (History)

${PAI_DIR}/history/research/YYYY-MM/YYYY-MM-DD_[topic]/
├── README.md
├── research-report.md
└── metadata.json

Key Principles

1. Parallel execution - Launch multiple agents simultaneously
2. Hard timeouts - Don't wait indefinitely, proceed with partial results
3. Simplest first - Always try free tools before paid services
4. Auto-routing - Skill analyzes intent and activates appropriate workflow

WebSearch Tool Usage

Built-in Web Search (FREE)

Claude Code includes a built-in WebSearch tool for real-time web queries.

When to Use:

• Current events and recent information
• Documentation and API references
• Pricing, availability, status checks
• Fact verification beyond training data

Best Practices:

// Include year for recent info
WebSearch({ query: "Next.js 15 features 2024" })

// Be specific
WebSearch({ query: "TypeScript 5.4 satisfies operator examples" })

// Use domain filtering for trusted sources
WebSearch({
  query: "React hooks best practices",
  allowed_domains: ["react.dev", "kentcdodds.com"]
})

Query Optimization:

• Include year: "React Server Components 2024"
• Be specific: "Bun vs Node.js benchmark comparison"
• Use domain filters: Focus on official docs or trusted sources

Source Citation: Always cite sources in research output:

## Sources
- [React Documentation](https://react.dev/...)
- [Official Blog Post](https://...)

WebSearch vs WebFetch

Integration with Research Workflows

WebSearch is the foundation of claude-researcher agent:

1. Query decomposition into sub-queries
2. Parallel WebSearch calls
3. Result synthesis
4. Source attribution

Workflow Files