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.
The simplest way to conduct deep research:
/deep-research [your research topic]
Example:
/deep-research AI applications in clinical diagnosis
This command will:
- Refine your research question (ask clarifying questions)
- Create a research plan
- Deploy multiple research agents in parallel
- Synthesize findings into a comprehensive report
- Validate all citations
- Output results to
RESEARCH/[topic]/directory
For more control, use commands sequentially:
/refine-question [your raw question]
The Question Refiner will:
- Ask 5-6 clarifying questions
- Understand your specific needs
- Generate a structured research prompt
/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
/deep-research [your topic]
Or use the research-executor skill directly with the structured prompt.
/synthesize-findings RESEARCH/[topic]/research_notes/
/validate-citations RESEARCH/[topic]/full_report.md
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/
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:
- Asks clarifying questions about:
- Specific focus areas
- Output format requirements
- Geographic and time scope
- Target audience
- Special requirements
- 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/
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:
- Phase 1: Verifies question scoping (already done by Question Refiner)
- Phase 2: Creates retrieval plan with subtopics
- Phase 3: Deploys multiple parallel research agents
- Phase 4: Triangulates sources (cross-validation)
- Phase 5: Synthesizes knowledge
- Phase 6: Performs quality assurance
- Phase 7: Generates formatted output
File: .claude/skills/research-executor/
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
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:
- Checks every factual claim has a citation
- Verifies citation completeness (author, date, title, URL, pages)
- Rates source quality (A-E scale)
- Verifies citations support the claims
- Detects hallucinations
- 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/
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:
- Groups findings by themes
- Identifies consensus and contradictions
- Resolves conflicting information
- Creates logical narrative flow
- Extracts actionable insights
- Identifies research gaps
File: .claude/skills/synthesizer/
Usage: /deep-research [research topic]
Description: Execute complete deep research workflow
Workflow:
- Question Refiner (ask clarifying questions)
- Research Executor (7-phase process)
- Synthesizer (combine findings)
- Citation Validator (verify quality)
Output: Complete research package in RESEARCH/[topic]/
Usage: /refine-question [raw question]
Description: Refine question into structured prompt
Output: Structured research prompt ready for research
Usage: /plan-research [structured prompt]
Description: Create detailed research execution plan
Output: Research plan with subtopics and agent deployment
Usage: /validate-citations [file path]
Description: Validate citations in research document
Output: Validation report with issues and corrections
Usage: /synthesize-findings [directory path]
Description: Synthesize multiple research findings
Output: Unified research report
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/
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]
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]
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
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"
The Question Refiner asks questions for a reason. Detailed answers lead to better research.
When using /plan-research, review the plan before execution. Adjust if needed.
Always use /validate-citations on final reports to ensure quality.
Read executive_summary.md first, then dive into full_report.md for details.
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"
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]
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]"
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)
Solution: The synthesizer skill handles this. If overwhelmed:
- Focus on executive summary first
- Read specific sections of full report
- Check research_notes for raw agent outputs
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
Solution: This is normal. The synthesizer will:
- Acknowledge contradictions
- Explain why sources might disagree
- Present multiple perspectives
- Note uncertainty where it exists
Use deep research to gather information before writing:
/deep-research [topic]- Review
executive_summary.mdandfull_report.md - Use findings as foundation for your writing
- Include citations from bibliography
Use deep research to inform decisions:
/deep-research [decision topic]- Focus on "Recommendations" section
- Review "Contradictions" to understand risks
- Check "Limitations" for unknowns
Use deep research for literature reviews:
- Specify "peer-reviewed sources only"
- Use citation validator to ensure quality
- Review bibliography for relevant papers
- Build on identified research gaps
- 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
- Be specific: Narrower topics = faster research
- Limit timeframes: "Last 2 years" faster than "Last 10 years"
- Limit geography: "US market" faster than "Global"
- Specify source types: "Industry reports only" faster than "All sources"
- Use GoT for complex topics: Higher quality through strategic exploration
- Request citation validation: Ensures accuracy
- Specify source quality requirements: "A-B rated sources only"
- Allow more time: Quality research takes time
-
Additional Skills:
- Domain-specific refinements (healthcare.md, financial.md, legal.md)
- Translation and localization
- Automated presentation generation
-
Enhanced GoT Operations:
- Backtracking operations
- Dynamic frontier management
- Parallel aggregation
-
Integration Options:
- Export to different formats (PDF, PowerPoint)
- API access for programmatic research
- Web interface for non-technical users
To contribute new skills or improvements:
- Follow the skill structure in
.claude/skills/ - Include skill.json, instructions.md, examples.md
- Test with diverse research topics
- Document usage in this guide
- 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
- 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.