Skip to content

Latest commit

 

History

History
593 lines (431 loc) · 17 KB

File metadata and controls

593 lines (431 loc) · 17 KB

Claude Code Deep Research Agent - Implementation Guide

Overview

This guide explains how to use the Claude Code Deep Research Agent framework to conduct comprehensive, multi-agent research with proper citations and high-quality outputs.

Quick Start

1. Basic Usage - Single Command

The simplest way to conduct deep research:

/deep-research [your research topic]

Example:

/deep-research AI applications in clinical diagnosis

This command will:

  1. Refine your research question (ask clarifying questions)
  2. Create a research plan
  3. Deploy multiple research agents in parallel
  4. Synthesize findings into a comprehensive report
  5. Validate all citations
  6. Output results to RESEARCH/[topic]/ directory

2. Step-by-Step Usage

For more control, use commands sequentially:

Step 1: Refine Your Question

/refine-question [your raw question]

The Question Refiner will:

  • Ask 5-6 clarifying questions
  • Understand your specific needs
  • Generate a structured research prompt

Step 2: Plan Research (Optional)

/plan-research [structured prompt from Step 1]

This creates a detailed research execution plan showing:

  • How the topic will be broken into subtopics
  • Which agents will be deployed
  • Expected timeline

Step 3: Execute Research

/deep-research [your topic]

Or use the research-executor skill directly with the structured prompt.

Step 4: Synthesize Findings (If needed)

/synthesize-findings RESEARCH/[topic]/research_notes/

Step 5: Validate Citations

/validate-citations RESEARCH/[topic]/full_report.md

Project Structure

claude-code-deep-research/
├── .claude/
│   ├── skills/                    # Skills definitions
│   │   ├── question-refiner/      # Question refinement skill
│   │   │   ├── skill.json
│   │   │   ├── instructions.md
│   │   │   └── examples.md
│   │   ├── research-executor/     # Main research execution skill
│   │   │   ├── skill.json
│   │   │   ├── instructions.md
│   │   │   └── examples.md
│   │   ├── got-controller/        # Graph of Thoughts controller
│   │   │   ├── skill.json
│   │   │   ├── instructions.md
│   │   │   └── examples.md
│   │   ├── citation-validator/    # Citation validation skill
│   │   │   ├── skill.json
│   │   │   ├── instructions.md
│   │   │   └── examples.md
│   │   └── synthesizer/           # Research synthesis skill
│   │       ├── skill.json
│   │       ├── instructions.md
│   │       └── examples.md
│   └── commands/                  # User-facing commands
│       ├── deep-research/         # Main deep research command
│       ├── refine-question/       # Question refinement command
│       ├── plan-research/         # Research planning command
│       ├── validate-citations/    # Citation validation command
│       └── synthesize-findings/   # Synthesis command
├── CLAUDE.md                      # Main documentation
├── CLAUDE2.md                     # GoT framework documentation
├── PROJECT_UNDERSTANDING.md       # Project understanding
├── IMPLEMENTATION_GUIDE.md        # This file
├── deepresearchprocess.md         # 7-phase research process
└── RESEARCH/                      # Research outputs (created during use)
    └── [topic_name]/
        ├── README.md
        ├── executive_summary.md
        ├── full_report.md
        ├── data/
        ├── visuals/
        ├── sources/
        ├── research_notes/
        └── appendices/

Skills Reference

Question Refiner

Purpose: Transform raw research questions into structured research prompts

When to Use:

  • You have a vague topic or question
  • You need help defining research scope
  • You want to clarify output requirements

What it Does:

  1. Asks clarifying questions about:
    • Specific focus areas
    • Output format requirements
    • Geographic and time scope
    • Target audience
    • Special requirements
  2. Generates structured prompt with:
    • TASK (clear objective)
    • CONTEXT/BACKGROUND (why this matters)
    • SPECIFIC QUESTIONS (3-7 concrete sub-questions)
    • KEYWORDS (search terms)
    • CONSTRAINTS (time, geography, source types)
    • OUTPUT FORMAT (detailed structure)

File: .claude/skills/question-refiner/

Research Executor

Purpose: Execute the full 7-phase deep research process

When to Use:

  • You have a structured research prompt
  • You want comprehensive research with proper citations
  • You need multi-agent parallel research

What it Does:

  1. Phase 1: Verifies question scoping (already done by Question Refiner)
  2. Phase 2: Creates retrieval plan with subtopics
  3. Phase 3: Deploys multiple parallel research agents
  4. Phase 4: Triangulates sources (cross-validation)
  5. Phase 5: Synthesizes knowledge
  6. Phase 6: Performs quality assurance
  7. Phase 7: Generates formatted output

File: .claude/skills/research-executor/

GoT Controller

Purpose: Manage research as a Graph of Thoughts for complex topics

When to Use:

  • Research topic is complex or multifaceted
  • You need strategic exploration (depth vs breadth)
  • High-stakes research where quality is critical

What it Does:

  • Maintains graph state (nodes, edges, scores)
  • Executes GoT operations:
    • Generate(k): Spawn k parallel research paths
    • Aggregate(k): Combine k findings
    • Refine(1): Improve existing finding
    • Score: Rate quality (0-10)
    • KeepBestN(n): Prune to top n nodes
  • Optimizes research quality through strategic operations

File: .claude/skills/got-controller/

Patterns:

  • Balanced: Generate(4-5) → Score best → Deepen top paths → Aggregate
  • Depth-first: Generate(3) → Take best → Generate(3) from it → Deep dive
  • Breadth-first: Generate(8) → KeepBestN(5) → Generate(2) from each → Aggregate

Citation Validator

Purpose: Ensure all claims have accurate, complete, high-quality citations

When to Use:

  • Finalizing a research report
  • Reviewing someone else's research
  • Before publishing or sharing research

What it Does:

  1. Checks every factual claim has a citation
  2. Verifies citation completeness (author, date, title, URL, pages)
  3. Rates source quality (A-E scale)
  4. Verifies citations support the claims
  5. Detects hallucinations
  6. Provides correction recommendations

Quality Scale:

  • A: Peer-reviewed, systematic reviews, RCTs
  • B: Cohort studies, clinical guidelines, reputable analysts
  • C: Expert opinion, case reports, mechanistic studies
  • D: Preprints, preliminary research, blogs
  • E: Anecdotal, theoretical, speculative

File: .claude/skills/citation-validator/

Synthesizer

Purpose: Combine multiple research findings into coherent reports

When to Use:

  • Multiple agents have completed research
  • You need to combine findings into unified report
  • There are contradictions between sources

What it Does:

  1. Groups findings by themes
  2. Identifies consensus and contradictions
  3. Resolves conflicting information
  4. Creates logical narrative flow
  5. Extracts actionable insights
  6. Identifies research gaps

File: .claude/skills/synthesizer/

Commands Reference

/deep-research

Usage: /deep-research [research topic]

Description: Execute complete deep research workflow

Workflow:

  1. Question Refiner (ask clarifying questions)
  2. Research Executor (7-phase process)
  3. Synthesizer (combine findings)
  4. Citation Validator (verify quality)

Output: Complete research package in RESEARCH/[topic]/

/refine-question

Usage: /refine-question [raw question]

Description: Refine question into structured prompt

Output: Structured research prompt ready for research

/plan-research

Usage: /plan-research [structured prompt]

Description: Create detailed research execution plan

Output: Research plan with subtopics and agent deployment

/validate-citations

Usage: /validate-citations [file path]

Description: Validate citations in research document

Output: Validation report with issues and corrections

/synthesize-findings

Usage: /synthesize-findings [directory path]

Description: Synthesize multiple research findings

Output: Unified research report

Usage Examples

Example 1: Market Research

User: /deep-research AI in healthcare market

Claude: I'll help you research AI in healthcare. Let me ask some questions:

1. What specific aspects interest you most?
   - Clinical diagnosis and treatment
   - Drug discovery
   - Hospital operations
   - All of the above

2. What output format do you prefer?
   - Executive summary (3-5 pages)
   - Comprehensive report (20-30 pages)
   - Market analysis with projections

[... more questions ...]

User: Focus on clinical diagnosis, comprehensive report, last 3 years data,
    global scope, audience is healthcare executives

Claude: Perfect. I'll create a research plan and execute multi-agent research.

[Deploys 6 parallel research agents...]

[Synthesizes findings...]

[Validates citations...]

Research complete! Output in RESEARCH/AI_Clinical_Diagnosis_2024/

Example 2: Technical Assessment

User: /refine-question Should I use WebAssembly for my project?

Claude: I'll help you research WebAssembly. Quick questions:

1. What type of project?
   - Web application
   - Game development
   - Data processing
   - Other

2. What's your main concern?
   - Performance
   - Browser compatibility
   - Development complexity
   - All of the above

[... generates structured research prompt ...]

User: /deep-research [uses the structured prompt]

[Research executes with technical focus]

Example 3: Academic Literature Review

User: /deep-research Transformer architectures in AI, focus on efficiency improvements

Claude: [Asks academic-level questions]

User: Peer-reviewed sources only, 2017-present, comprehensive literature review

Claude: [Deploys academic research agents focusing on:
      - Seminal papers
      - Efficiency techniques (distillation, quantization)
      - Recent advances (2022-2024)
      - Theoretical analyses]

[Generates academic literature review with proper citations]

Research Output Structure

Every /deep-research command creates a structured output:

RESEARCH/[topic_name]/
├── README.md                    # Overview and navigation
├── executive_summary.md         # 1-2 page key findings
├── full_report.md               # Complete research report
├── data/
│   ├── statistics.md            # Key numbers and facts
│   └── key_facts.md             # Important findings
├── visuals/
│   └── descriptions.md          # Chart/graph descriptions
├── sources/
│   ├── bibliography.md          # Complete citations
│   └── source_quality_table.md  # A-E quality ratings
├── research_notes/
│   └── agent_findings_summary.md # Raw agent outputs
└── appendices/
    ├── methodology.md           # Research methods
    └── limitations.md           # What couldn't be determined

Best Practices

1. Be Specific with Your Research Question

Bad: "Research AI"

Good: "Research AI applications in clinical diagnosis and treatment for healthcare executives making adoption decisions, focusing on US and European markets, 2022-2024 data with projections to 2028"

2. Answer Clarifying Questions Thoroughly

The Question Refiner asks questions for a reason. Detailed answers lead to better research.

3. Review the Research Plan

When using /plan-research, review the plan before execution. Adjust if needed.

4. Validate Citations Before Publishing

Always use /validate-citations on final reports to ensure quality.

5. Start with Executive Summary

Read executive_summary.md first, then dive into full_report.md for details.

Advanced Usage

Using Specific Skills Directly

You can invoke skills without commands:

"Use the question-refiner skill to help refine my research question about [topic]"
"Use the got-controller skill to manage this research using balanced exploration pattern"

Custom Research Patterns

For comprehensive research:

1. /refine-question [topic]
2. /plan-research [structured prompt]
3. Review and adjust plan
4. Execute research (agents will be deployed)
5. /synthesize-findings [research_notes/]
6. /validate-citations [full_report.md]

For quick research:

/deep-research [specific, well-defined topic]

Combining with GoT Operations

For complex research, explicitly request GoT patterns:

"Use the got-controller with depth-first exploration for research on [technical topic]"
"Use the got-controller with breadth-first exploration for [trend analysis topic]"

Troubleshooting

Issue: Research takes too long

Solution: Narrow the scope. Be more specific about:

  • Geographic focus (one country instead of global)
  • Time period (last 2 years instead of last 10)
  • Specific aspects (one application instead of all applications)

Issue: Too many results, hard to synthesize

Solution: The synthesizer skill handles this. If overwhelmed:

  1. Focus on executive summary first
  2. Read specific sections of full report
  3. Check research_notes for raw agent outputs

Issue: Citations fail validation

Solution: Common citation problems:

  • Missing URLs → Use WebSearch to find sources
  • Incomplete citations → Add author, date, title
  • Dead links → Use archive.org or find alternatives
  • Run citation validator again after corrections

Issue: Contradictory findings

Solution: This is normal. The synthesizer will:

  1. Acknowledge contradictions
  2. Explain why sources might disagree
  3. Present multiple perspectives
  4. Note uncertainty where it exists

Integration with Existing Workflows

As a Pre-writing Step

Use deep research to gather information before writing:

  1. /deep-research [topic]
  2. Review executive_summary.md and full_report.md
  3. Use findings as foundation for your writing
  4. Include citations from bibliography

For Decision Making

Use deep research to inform decisions:

  1. /deep-research [decision topic]
  2. Focus on "Recommendations" section
  3. Review "Contradictions" to understand risks
  4. Check "Limitations" for unknowns

For Academic Work

Use deep research for literature reviews:

  1. Specify "peer-reviewed sources only"
  2. Use citation validator to ensure quality
  3. Review bibliography for relevant papers
  4. Build on identified research gaps

Performance Considerations

Research Time Estimates

  • Quick research (narrow topic): 10-15 minutes
  • Standard research (moderate scope): 20-30 minutes
  • Comprehensive research (broad scope): 30-60 minutes
  • Academic literature review: 45-90 minutes

Factors affecting time:

  • Number of subtopics (3-7)
  • Number of agents deployed (3-8)
  • Source availability
  • Web search speed

Optimizing for Speed

  1. Be specific: Narrower topics = faster research
  2. Limit timeframes: "Last 2 years" faster than "Last 10 years"
  3. Limit geography: "US market" faster than "Global"
  4. Specify source types: "Industry reports only" faster than "All sources"

Optimizing for Quality

  1. Use GoT for complex topics: Higher quality through strategic exploration
  2. Request citation validation: Ensures accuracy
  3. Specify source quality requirements: "A-B rated sources only"
  4. Allow more time: Quality research takes time

Future Enhancements

Planned Features

  1. Additional Skills:

    • Domain-specific refinements (healthcare.md, financial.md, legal.md)
    • Translation and localization
    • Automated presentation generation
  2. Enhanced GoT Operations:

    • Backtracking operations
    • Dynamic frontier management
    • Parallel aggregation
  3. Integration Options:

    • Export to different formats (PDF, PowerPoint)
    • API access for programmatic research
    • Web interface for non-technical users

Contributing

To contribute new skills or improvements:

  1. Follow the skill structure in .claude/skills/
  2. Include skill.json, instructions.md, examples.md
  3. Test with diverse research topics
  4. Document usage in this guide

Support and Documentation

  • CLAUDE.md: Main documentation with 7-phase process
  • CLAUDE2.md: Graph of Thoughts framework details
  • PROJECT_UNDERSTANDING.md: Project architecture and design
  • deepresearchprocess.md: 7-phase deep research methodology
  • Skill examples: Each skill has examples.md with detailed examples

Version History

  • v1.0 (2024-12-25): Initial implementation
    • Question Refiner skill
    • Research Executor skill
    • GoT Controller skill
    • Citation Validator skill
    • Synthesizer skill
    • 5 commands for common workflows

For questions or issues, refer to the skill-specific examples.md files or the main CLAUDE.md documentation.