Comprehensive Deep Research Orchestration (Tier 4 RBMAS)

Guide Main Claude to coordinate comprehensive multi-phase research using 7-phase RBMAS methodology for established domains requiring 4+ dimensions of analysis.

Quick Start

1. Analyze query dimensions - Identify 4+ subtopics/angles requiring deep investigation
2. Setup progress tracking - Use TodoWrite to track 7 RBMAS phases
3. Execute RBMAS phases - Follow structured methodology with quality gates
4. Spawn specialist subagents - Minimum 3-5 agents (web-researcher, academic-researcher, fact-checker)
5. Validate quality - Check citation density, source diversity, gap detection
6. Synthesize findings - Combine results into comprehensive report
7. Report completion - Deliver comprehensive analysis with citations

7-Phase RBMAS Methodology

Phase 1: SCOPE

Objective: Break down research question into focused sub-questions

MUST:

• Analyze the research question thoroughly
• Identify 5-7 key concepts and sub-questions
• Define research boundaries and objectives
• Map dimensions to specific investigation areas

Example sub-questions for "WebRTC security":

• Cryptographic protocols (DTLS, SRTP)
• Network security (ICE, TURN vulnerabilities)
• Implementation security (browser implementations)
• Authentication mechanisms
• Attack vectors and mitigations

Phase 2: PLAN

Objective: Develop comprehensive research strategy

MUST:

• Create strategy for each sub-question
• Identify required sources (web, academic, documentation)
• Plan verification and cross-referencing approach
• Estimate depth and breadth needed per dimension
• Allocate specialist agents to dimensions

Phase 3: RETRIEVE (MANDATORY SPAWNING)

Objective: Gather information through specialist subagents

🚨 CRITICAL: Main Claude MUST spawn subagents. NEVER skip this phase.

MUST:

• Spawn 3-7 specialist subagents using Task tool
• Use web-researcher for general web queries
• Use academic-researcher for scholarly sources
• Use search-specialist for complex Boolean queries
• Spawn agents IN PARALLEL (not sequential)
• Pass all research parameters to subagents

Example subagent spawn:

Task(
  subagent_type: "web-researcher",
  description: "Research WebRTC cryptographic protocols",
  prompt: "Research DTLS and SRTP protocols used in WebRTC...

Focus:
- Protocol specifications
- Implementation details
- Security considerations
- Current best practices (2024-2025)

Requirements:
- Multiple authoritative sources
- Include URLs and timestamps
- Cite specifications and standards",
  model: "sonnet"
)

Phase 4: TRIANGULATE (MANDATORY VERIFICATION)

Objective: Cross-reference and verify findings

🚨 CRITICAL: Main Claude MUST spawn fact-checker for critical claims.

MUST:

• Cross-reference findings across sources
• Identify consensus vs conflicting information
• Spawn fact-checker agent for critical claims
• Flag uncertainties and knowledge gaps
• Validate citations and source quality

Quality Gates:

• ✅ Citation density: Minimum 3 sources per major claim
• ✅ Source diversity: Mix of academic, industry, official docs
• ✅ Gap detection: Explicitly flag missing information

Phase 5: DRAFT

Objective: Synthesize findings into coherent narrative

MUST:

• Synthesize findings from all subagents
• Organize by themes and importance
• Include evidence and citations inline
• Maintain objectivity and acknowledge limitations
• Structure report with clear sections

Phase 6: CRITIQUE

Objective: Review for quality and completeness

MUST:

• Review for accuracy, completeness, and bias
• Identify gaps requiring additional research
• Verify citation accuracy
• Ensure claims properly supported
• Check against quality gates

Phase 7: PACKAGE

Objective: Deliver final comprehensive report

MUST:

• Format final research report in markdown
• Include executive summary and bibliography
• Add confidence assessments per claim
• Document methodology and sources
• Save comprehensive report

Delegation Rules (CRITICAL)

Main Claude MUST:

1. ❌ NEVER research directly - Always spawn subagents for information gathering
2. ✅ ALWAYS spawn 3-7 specialist subagents (one or more per dimension)
3. ✅ ALWAYS spawn fact-checker in Phase 4 for verification
4. ✅ SPAWN agents IN PARALLEL (not sequential) for efficiency
5. ✅ PASS research parameters to all subagents clearly

Minimum subagent requirements:

• 3-5 researchers (web-researcher, academic-researcher, search-specialist)
• 1 fact-checker (verification)
• Optional: synthesis-researcher (if not doing synthesis yourself)

Quality Gates (Mandatory Validation)

Before completing Phase 7, VERIFY:

✅ Citation Density:

• Minimum 3 authoritative sources per major claim
• Sources include URLs and timestamps
• Mix of academic, industry, official documentation

✅ Source Diversity:

• NOT all sources from single domain
• Multiple independent authorities
• Recent sources (2023-2025 preferred)

✅ Gap Detection:

• Explicitly flag missing information
• Acknowledge conflicting data
• Note areas requiring further research

✅ Methodology Documentation:

• 7 phases all executed
• Subagents spawned as required
• Verification performed

Examples

Example 1: Comprehensive 4-Dimension Query

Query: "Research WebRTC security across cryptography, network, implementation, and browser dimensions"

Response:

I'll coordinate comprehensive WebRTC security research using 7-phase RBMAS methodology.

[TodoWrite: Tracks 7 phases]

Phase 1 (SCOPE): Breaking into 4 dimensions:
1. Cryptographic protocols (DTLS, SRTP)
2. Network security (ICE, TURN, STUN)
3. Implementation security (codebase analysis)
4. Browser-specific security (Chrome, Firefox, Safari)

Phase 2 (PLAN): Allocating specialists:
- academic-researcher: Cryptographic protocol research
- web-researcher: Network security and browser implementations
- search-specialist: Implementation vulnerabilities
- fact-checker: Verify critical security claims

Phase 3 (RETRIEVE): Spawning subagents...
[Spawns 4 agents IN PARALLEL using Task tool]

[Wait for agents to complete]

Phase 4 (TRIANGULATE): Verifying findings...
[Spawns fact-checker for critical claims]

Phases 5-7: Drafting, critiquing, packaging comprehensive report...

[Delivers 4-dimension comprehensive analysis with citations]

Example 2: BAD - Skip Spawning (VIOLATION)

Query: "Research cloud gaming latency optimization comprehensively"

❌ WRONG Response:

Let me research cloud gaming latency optimization.
[Directly uses WebSearch without spawning subagents]

Why Wrong: Tier 4 MUST orchestrate subagents through 7 phases, NOT research directly.

✅ CORRECT Response:

I'll coordinate comprehensive cloud gaming latency research using RBMAS.

Phase 1 (SCOPE): 5 dimensions identified...
Phase 2 (PLAN): Strategy developed...
Phase 3 (RETRIEVE): Spawning specialists...
[Spawns web-researcher, academic-researcher, search-specialist]

Example 3: Quality Gate Failure → Retry

Scenario: Phase 6 CRITIQUE detects gaps

Response:

Phase 6 (CRITIQUE): Quality gate check...

⚠️ Gap detected: Insufficient sources on browser-specific implementations
- Only 1 source for Firefox implementation
- No sources for Safari implementation

Returning to Phase 3 (RETRIEVE):
[Spawns additional search-specialist for browser implementations]

[After retrieval] Phase 6 (CRITIQUE): Re-checking quality gates...
✅ Citation density: Passed (3+ sources per claim)
✅ Source diversity: Passed (academic + industry + official docs)
✅ Gap detection: Passed (all dimensions covered)

Proceeding to Phase 7 (PACKAGE)...

Response Style

Main Claude MUST:

• Be concise and direct (2-4 sentences per phase update)
• Use TodoWrite to track phase progress
• Report subagent spawn status clearly
• Indicate quality gate results explicitly
• Deliver comprehensive final synthesis

Main Claude MUST NOT:

• Skip phases or subagent spawning
• Research directly instead of delegating
• Ignore quality gates
• Provide superficial analysis

Key Principles

1. Depth over Speed: Take time to explore thoroughly across all dimensions
2. Source Diversity: Use multiple independent authoritative sources
3. Transparency: Cite sources, acknowledge limitations, flag gaps
4. Objectivity: Present facts, note when conclusions are inferential
5. Iteration: Refine understanding through quality gates and re-retrieval
6. Orchestration: Coordinate multiple specialists, never research alone

When to Use This Skill

✅ USE internet-deep-orchestrator when:

• Query has 4+ distinct dimensions or subtopics
• Comprehensive analysis required (not simple lookup)
• Multiple specialist perspectives needed
• Quality validation critical (citations, verification)
• Iterative refinement valuable

❌ DON'T USE (use other tiers instead):

• Simple lookups (Tier 1: web-researcher)
• Focused single-topic research (Tier 2: specialist agents)
• 2-3 dimension queries (Tier 3: internet-light-orchestrator)
• Novel/emerging domains (Tier 5: internet-research-orchestrator)

Success Criteria

Phase 3 complete when:

• ✅ Minimum 3-5 subagents spawned
• ✅ All dimensions covered by specialists
• ✅ Agents spawned in parallel (not sequential)

Phase 7 complete when:

• ✅ All 7 phases executed
• ✅ Quality gates passed (citations, diversity, gaps)
• ✅ Comprehensive report delivered with bibliography
• ✅ Confidence assessments provided per claim

Methodology: RBMAS (Research-Based Multi-Agent System) Complexity: Tier 4 (Comprehensive) Dimensions: 4+ subtopics Subagents: 3-7 specialists (minimum) Quality Gates: Mandatory (citation density, source diversity, gap detection)