majiayu000 Ultimate Research - Elite Multi-Agent Research Orchestration
Core Capability
This skill orchestrates the most comprehensive research capability available by:
- Parallel agent execution: Simultaneously deploy all 4 specialized research agents
- Full MCP utilization: Leverage every tool from ref, exa, brave-search, and context7 MCPs
- Best practices adherence: Follow MCP_BEST_PRACTICES_COMPREHENSIVE_GUIDE.md for optimal performance
- Sequential thinking: Use complex reasoning with sequential-thinking MCP
- Token efficiency: 96% reduction via smart routing and caching
When This Skill Activates
Automatically triggers during /sc:research commands or when:
- User requests comprehensive research
- Multi-dimensional analysis needed
- Current events + historical context required
- Code examples + documentation + community insights needed
- Competitive analysis requiring multiple perspectives
- Academic/technical research with citation requirements
Research Architecture
Tier 1: Parallel Agent Deployment
Execute simultaneously for maximum efficiency:
┌─────────────────────────────────────────┐
│ Research Query Received │
└──────────────┬──────────────────────────┘
│
┌─────────┼──────────┬──────────┐
│ │ │ │
▼ ▼ ▼ ▼
┌─────────┐ ┌──────┐ ┌────────┐ ┌────────┐
│ deep- │ │search│ │ trend- │ │ ux- │
│research │ │spec │ │research│ │research│
│ -agent │ │ │ │ -er │ │ -er │
└─────────┘ └──────┘ └────────┘ └────────┘
│ │ │ │
└─────────┴──────────┴──────────┘
│
▼
┌────────────────────┐
│ Synthesis & │
│ Validation │
└────────────────────┘
Agent Responsibilities:
- deep-research-agent: Complex multi-hop analysis, evidence chains, synthesis
- search-specialist: Web search, documentation discovery, source verification
- trend-researcher: Market trends, viral content, emerging patterns
- ux-researcher: User behavior, pain points, feedback analysis
Tier 2: MCP Tool Orchestration
Use decision tree based on MCP_BEST_PRACTICES_COMPREHENSIVE_GUIDE.md:
Documentation Queries:
Primary: ref MCP → Fast, token-efficient (96% reduction)
Fallback: exa MCP → Code context if docs insufficient
Tertiary: brave-search MCP → General web if needed
Code Example Queries:
Primary: exa MCP (get_code_context_exa) → Semantic search, fastest (1.18s)
Fallback: ref MCP → Official documentation
Tertiary: brave-search MCP → Tutorials and guides
News/Current Events:
Primary: brave-search MCP (news_search, freshness='pd') → Latest information
Secondary: exa MCP → Deep analysis
Verification: context7 MCP → Documentation updates
Research/Semantic:
Primary: exa MCP (web_search_exa) → AI-native search
Secondary: brave-search MCP → Broad coverage
Documentation: ref MCP + context7 MCP → Technical verification
Tier 3: Sequential Thinking Integration
For complex reasoning tasks, invoke sequential-thinking MCP:
When to use:
- Multi-step logical deduction needed
- Conflicting information requires resolution
- Hypothesis generation and validation
- Cost-benefit analysis across MCPs
- Quality vs speed tradeoffs
Execution Protocol
Phase 1: Parallel Agent Launch (0-30 seconds)
Launch all 4 agents simultaneously in single message:
[Task tool invocation with 4 parallel calls]
deep-research-agent task: [research scope]
search-specialist task: [documentation discovery]
trend-researcher task: [market analysis]
ux-researcher task: [user insights]
Critical: Use single message with multiple Task tool calls for true parallelism
Phase 2: MCP Tool Utilization (30s - 5min)
Based on agent requirements, execute MCP tools following best practices:
Token Optimization:
- ref MCP: Use for all documentation (automatic 5k token limit)
- exa MCP: Configure tokensNum=8000 for code research
- brave-search MCP: Set count=10, result_filter=["web","news"]
- context7 MCP: Use mode='code' for API refs, mode='info' for guides
Parallel Execution Example:
# Execute independent searches simultaneously
await asyncio.gather(
ref_search_documentation("React hooks TypeScript"),
exa_code_context("React hooks examples", tokensNum=8000),
brave_web_search("React hooks best practices 2024"),
context7_get_library_docs("/facebook/react", topic="hooks")
)
Caching Strategy (from MCP best practices):
Documentation (ref): 7-day TTL
Code examples (exa): 1-day TTL
News (brave): 1-hour TTL
General web: 6-hour TTL
Phase 3: Sequential Thinking (as needed)
When complex reasoning required:
Use sequential-thinking MCP to:
1. Analyze conflicting sources
2. Score result relevance
3. Identify information gaps
4. Generate synthesis strategy
5. Validate conclusions
Phase 4: Synthesis & Validation (final 20%)
Merge findings from all agents and MCPs:
1. Deduplicate results (by URL/content hash)
2. Score relevance (0-100 scale)
3. Rank by authority + freshness + completeness
4. Cross-verify claims across sources
5. Generate confidence scores
6. Build evidence chains
MCP Best Practices Implementation
From MCP_BEST_PRACTICES_COMPREHENSIVE_GUIDE.md
Performance Optimization:
- Enable only needed Exa tools:
--tools=get_code_context_exa,web_search_exa - Use Ref HTTP method (fastest setup)
- Implement request queuing for rate limits
- Progressive loading: summaries first, details on-demand
Cost Optimization:
- Budget-aware routing (see guide Section 7.3)
- Aggressive caching (60-80% hit rate achievable)
- Token limit tuning per query type
- Free tier maximization strategy
Quality Optimization:
- Result scoring algorithm (relevance + freshness + authority + completeness)
- Confidence thresholds by query type
- Fallback chains for reliability
- Multi-MCP verification for critical info
Output Format
Research Report Structure
# [Research Topic]
**Research Date**: YYYY-MM-DD
**Confidence Score**: XX/100
**Sources Consulted**: N unique sources
## Executive Summary
[3-5 sentence overview of key findings]
## Key Findings
### 1. [Finding Category]
- **Source**: [Agent + MCP used]
- **Evidence**: [Supporting data]
- **Confidence**: [High/Medium/Low]
- **Citations**: [URLs/references]
### 2. [Finding Category]
...
## Detailed Analysis
### [Section 1]: Official Documentation
**Source**: ref MCP + context7 MCP
[Findings from authoritative sources]
### [Section 2]: Code Examples & Implementations
**Source**: exa MCP
[Real-world usage patterns]
### [Section 3]: Community Insights
**Source**: brave-search MCP + search-specialist agent
[Developer feedback, common issues]
### [Section 4]: Market Trends
**Source**: trend-researcher agent + exa MCP
[Current state, emerging patterns]
### [Section 5]: User Experience Analysis
**Source**: ux-researcher agent
[Pain points, opportunities]
## Synthesis & Recommendations
### Best Practices Identified
1. [Practice 1] - Source: [agents/MCPs]
2. [Practice 2] - Source: [agents/MCPs]
### Common Pitfalls
1. [Pitfall 1] - Source: [agents/MCPs]
2. [Pitfall 2] - Source: [agents/MCPs]
### Recommended Approach
[Synthesized recommendation based on all findings]
## Evidence Quality Assessment
| Category | Sources | Confidence | Notes |
|----------|---------|------------|-------|
| Official Docs | N | High | [ref/context7] |
| Code Examples | N | High | [exa] |
| Community | N | Medium | [brave/search-specialist] |
| Trends | N | Medium | [trend-researcher] |
| UX Insights | N | Low-Medium | [ux-researcher] |
## Information Gaps
[List any unanswered questions or areas needing deeper research]
## Sources
### Official Documentation
- [Source 1](URL) - ref MCP
- [Source 2](URL) - context7 MCP
### Code Repositories & Examples
- [Source 3](URL) - exa MCP
- [Source 4](URL) - exa MCP
### Community Discussions
- [Source 5](URL) - brave-search MCP
- [Source 6](URL) - search-specialist agent
### Market Analysis
- [Source 7](URL) - trend-researcher agent
- [Source 8](URL) - exa MCP
### User Research
- [Source 9](URL) - ux-researcher agent
- [Source 10](URL) - brave-search MCP
## Research Metadata
**Agents Used**:
- deep-research-agent: [time spent]
- search-specialist: [time spent]
- trend-researcher: [time spent]
- ux-researcher: [time spent]
**MCPs Used**:
- ref MCP: [queries count] - [token savings]
- exa MCP: [queries count] - [avg latency]
- brave-search MCP: [queries count] - [result types]
- context7 MCP: [queries count] - [mode used]
- sequential-thinking MCP: [thoughts generated]
**Performance Metrics**:
- Total research time: [duration]
- Parallel efficiency: [% time saved vs sequential]
- Cache hit rate: [%]
- Average query latency: [seconds]
- Total tokens consumed: [count]
- Token efficiency gain: [% vs baseline]
Advanced Patterns
Multi-Stage Research Workflow
For comprehensive research requiring multiple phases:
Stage 1: Discovery (Parallel agents + broad MCPs)
├─ Identify information landscape
├─ Map available sources
└─ Define research boundaries
Stage 2: Deep Dive (Focused MCP usage)
├─ Documentation analysis (ref + context7)
├─ Code pattern extraction (exa)
├─ Community sentiment (brave + search-specialist)
└─ Trend analysis (trend-researcher)
Stage 3: Verification (Cross-reference)
├─ Validate claims across sources
├─ Resolve contradictions (sequential-thinking)
└─ Assess confidence levels
Stage 4: Synthesis (deep-research-agent)
├─ Identify patterns
├─ Extract best practices
├─ Generate actionable insights
└─ Document evidence chains
Adaptive Depth Strategy
Match research depth to query importance:
Quick (5 min): Single-hop, primary MCPs only
- 1 agent (search-specialist)
- 2 MCPs (ref OR exa + brave)
- Summary output
Standard (15 min): Multi-hop, 2-3 agents
- 2-3 agents (search-specialist + deep-research-agent + trend-researcher)
- 3 MCPs (ref + exa + brave)
- Structured report
Deep (45 min): Comprehensive, all agents
- 4 agents (all parallel)
- 4 MCPs (ref + exa + brave + context7)
- Full report with evidence chains
Exhaustive (2+ hrs): Maximum depth, multi-stage
- 4 agents (parallel + sequential refinement)
- 4 MCPs + sequential-thinking
- Multi-stage workflow
- Iterative validation
- Complete documentation
Error Handling & Fallbacks
Rate Limiting
If MCP rate limited:
1. Implement exponential backoff (1s, 2s, 4s)
2. Switch to alternative MCP (ref → exa → brave)
3. Use cached results if available
4. Queue requests for later execution
Quality Threshold Violations
If results below confidence threshold:
1. Invoke fallback MCP chain
2. Cross-verify with additional agent
3. Use sequential-thinking to analyze gaps
4. Explicitly note uncertainty in output
Agent Failure
If agent fails to complete:
1. Continue with remaining agents
2. Note reduced coverage in metadata
3. Suggest re-running failed agent separately
4. Provide partial results with caveats
Performance Monitoring
Track these metrics per research session:
Efficiency Metrics:
- Parallel execution time savings
- Cache hit rate
- Token consumption vs baseline
- Cost per query
Quality Metrics:
- Source diversity (unique domains)
- Citation completeness
- Confidence score distribution
- Information gap rate
MCP Performance:
- Latency by MCP (ref: 1.7s, exa: 1.18s, brave: <2s)
- Success rate by MCP (target: >95%)
- Token efficiency (ref: 96% reduction)
- Fallback invocation frequency
Integration with /sc:research
This skill automatically activates when /sc:research is invoked. The command passes:
- Query string
- Depth flag (--depth)
- Strategy flag (--strategy)
- MCP flags (--ref, --exa, --brave-search, --c7)
- Thinking flags (--seq, --ultrathink)
The skill then:
- Parses flags to determine execution strategy
- Launches appropriate agents in parallel
- Orchestrates MCP tool usage per best practices
- Generates research report in standard format
- Saves to
claudedocs/research_[topic]_[timestamp].md
Examples
Example 1: Technical Documentation Research
Query: "Next.js App Router best practices and common pitfalls"
Execution:
1. Parallel agents launch (4 simultaneous)
2. ref MCP: Official Next.js docs (token-efficient)
3. exa MCP: GitHub repos with App Router examples
4. brave-search MCP: Reddit/GitHub discussions
5. context7 MCP: Vercel Next.js documentation
6. sequential-thinking: Analyze contradictions in community advice
7. Synthesis: Best practices + pitfalls + recommended approach
Output: Comprehensive guide with official docs + real code + community insights
Example 2: Market Research
Query: "AI coding assistant market analysis 2024"
Execution:
1. Parallel agents:
- search-specialist: Product websites, official announcements
- trend-researcher: TikTok/Twitter trends, viral content
- ux-researcher: User reviews, pain points
- deep-research-agent: Competitive analysis, feature comparison
2. brave-search MCP: Latest news, funding announcements
3. exa MCP: Company research, technical capabilities
4. ref MCP: API documentation analysis
5. sequential-thinking: Cost-benefit analysis, market positioning
Output: Market landscape + trends + competitive matrix + opportunities
Example 3: Technical Problem Solving
Query: "How to implement real-time collaboration in React with TypeScript"
Execution:
1. Parallel agents launch
2. ref MCP: React + TypeScript official docs
3. exa MCP:
- Code examples (WebSocket, CRDTs, Yjs)
- GitHub repos with implementations
4. brave-search MCP: Tutorials, blog posts, case studies
5. context7 MCP: Library docs (Socket.io, Y.js, etc.)
6. sequential-thinking: Evaluate approaches (WebSocket vs SSE vs WebRTC)
Output: Implementation guide + code examples + tradeoff analysis
Success Criteria
Research is successful when:
- All 4 agents executed (or failures documented)
- At least 3 MCPs utilized per best practices
- Parallel execution achieved (single message, multiple Task calls)
- Output includes confidence scores and source citations
- Evidence chains are traceable
- Information gaps explicitly identified
- Report saved to claudedocs/ directory
- Performance metrics documented
Troubleshooting
Issue: Agents not executing in parallel
Solution: Ensure single message with multiple Task tool calls, not sequential messages
Issue: High token consumption
Solution:
- Use ref MCP for docs (96% reduction)
- Configure exa tokensNum appropriately
- Enable only needed exa tools
- Implement caching
Issue: Rate limiting errors
Solution:
- Implement request queuing
- Use fallback MCP chains
- Enable exponential backoff
Issue: Low quality results
Solution:
- Increase depth level
- Add more agents
- Use sequential-thinking for gap analysis
- Enable multi-stage verification
References
- MCP Best Practices: See MCP_BEST_PRACTICES_COMPREHENSIVE_GUIDE.md
- Skill Best Practices: See skill.best.practices.md
- Research Command: See .claude/commands/sc/research.md
Skill Version: 1.0 Last Updated: 2025-11-24 Maintainer: Ultimate Research System License: Internal Use