deep-research（深入研究）

Deep Research Skill

Perform comprehensive technical research using the Firecrawl Agent API for autonomous web investigation. Use this skill when tasks require understanding external patterns, competitive analysis, or finding implementation examples.

When to Trigger Deep Research

Scan for these patterns in PRDs and task requirements:

Research Protocol

Step 1: Identify Research Needs

Before generating tasks, scan the PRD for:

1. External references - Named products, services, or standards
2. Comparative requirements - "better than", "like", "similar to"
3. Technical unknowns - Unfamiliar patterns or technologies
4. Best practice requests - "industry standard", "production-ready"

Step 2: Choose the Right Tool

Step 3: Execute Research

Using Firecrawl Agent

firecrawl_agent({
  prompt: "YOUR RESEARCH QUESTION - be specific",
  schema: {
    "type": "object",
    "properties": {
      "findings": {
        "type": "array",
        "items": {
          "type": "object",
          "properties": {
            "source": { "type": "string" },
            "approach": { "type": "string" },
            "details": { "type": "string" },
            "tradeoffs": { "type": "string" }
          }
        }
      },
      "recommendation": { "type": "string" }
    }
  }
})

Step 4: Structure Output

Always format research findings as:

## Research: [Topic]

### Summary
[2-3 sentence key takeaway]

### Findings

| Source | Approach | Key Details |
|--------|----------|-------------|
| Auth0 | JWT + refresh rotation | 15min access, 7d refresh |
| Clerk | Session tokens | Server-side validation |

### Recommendation
[How this applies to the current task]

### Sources
- [URL 1] - Description
- [URL 2] - Description

Common Research Patterns

Competitive Analysis

When PRD mentions competitors or "like X":

firecrawl_agent({
  prompt: "Compare how [Competitor A], [Competitor B], and [Competitor C] implement [feature]. Focus on [specific aspects from PRD].",
  schema: {
    "type": "object",
    "properties": {
      "providers": {
        "type": "array",
        "items": {
          "type": "object",
          "properties": {
            "name": { "type": "string" },
            "approach": { "type": "string" },
            "strengths": { "type": "string" },
            "weaknesses": { "type": "string" }
          }
        }
      },
      "recommendation": { "type": "string" }
    }
  }
})

Implementation Patterns

When PRD requires unfamiliar technology:

firecrawl_agent({
  prompt: "Find production examples of [technology] being used for [use case]. Include code patterns, gotchas, and performance considerations.",
  schema: {
    "type": "object",
    "properties": {
      "examples": {
        "type": "array",
        "items": {
          "type": "object",
          "properties": {
            "source": { "type": "string" },
            "pattern": { "type": "string" },
            "code_example": { "type": "string" },
            "gotchas": { "type": "string" }
          }
        }
      }
    }
  }
})

Architecture Research

When designing new systems:

firecrawl_agent({
  prompt: "What architectures do major [domain] platforms use for [requirement]? Compare approaches from [Company A], [Company B], etc.",
  schema: {
    "type": "object",
    "properties": {
      "architectures": {
        "type": "array",
        "items": {
          "type": "object",
          "properties": {
            "company": { "type": "string" },
            "architecture": { "type": "string" },
            "scale": { "type": "string" },
            "tradeoffs": { "type": "string" }
          }
        }
      }
    }
  }
})

Best Practices

When PRD requests "industry standard" approaches:

firecrawl_agent({
  prompt: "What are current best practices for [topic] in [year]? Focus on [specific requirements]. Include examples from production systems."
})

Integrating Research into Tasks

Research findings should be embedded in task details fields:

{
  "id": "5",
  "title": "Nova: Implement Refresh Token Rotation",
  "agentHint": "nova",
  "details": "## Requirements\nImplement refresh token rotation for session management.\n\n## Research Findings\nBased on competitive analysis:\n- Auth0: 15min access tokens, 7-day refresh tokens with rotation\n- Clerk: Session-based with server validation\n- Supabase: JWT with configurable expiry\n\n## Recommended Approach\nFollow Auth0 pattern with:\n- 15-minute access token lifetime\n- 7-day refresh token with single-use rotation\n- Revocation on suspicious activity\n\n## Code Signatures\n```typescript\nexport const refreshToken = Effect.gen(function* () {\n  // Implementation based on research\n})\n```"
}

Cost Management

Firecrawl Agent pricing is dynamic. Optimize costs:

1. Be specific - Vague prompts cost more
2. Use schemas - Structured output reduces processing
3. Provide URLs when known - Narrows search scope
4. Batch related questions - One comprehensive query vs multiple small ones

When NOT to Use Deep Research

• Library docs exist in Context7 - Use context7 instead
• You know the exact URL - Use firecrawl_scrape
• Simple factual lookup - Use firecrawl_search
• Code examples from GitHub repos - Use octocode_githubSearchCode (semantic search across repos)
• How React/major OSS projects do X - Use OctoCode to search their source

OctoCode Integration

For implementation pattern research, combine Firecrawl (web) with OctoCode (code):

# 1. Research how competitors approach the problem (web)
firecrawl_agent({ prompt: "How does Auth0 implement refresh token rotation?" })

# 2. Find actual implementations in open source (code)
octocode_githubSearchCode({
  query: "refresh token rotation",
  language: "typescript",
  stars: ">500"
})

# 3. Get library docs for the chosen approach
context7_get_library_docs({ libraryId: "/better-auth/better-auth", topic: "refresh tokens" })

Research Checklist

Before finalizing research-informed tasks:

• All "like X" and "similar to" references researched
• Competitive mentions analyzed
• Unfamiliar technologies investigated
• Research findings embedded in relevant task details
• Sources cited for verification
• Recommendations align with PRD requirements