shield_lock
JinDaGe
← Back to skills
extension
Category: AI Agent CapabilitiesNo API key required

create-research-brief

Two-phase research design and consolidation skill for multi-LLM optimized research

personAuthor: jakexiaohubgithub
open_in_newRepositorydownloadDownload package

Create Research Brief

A comprehensive two-phase skill for designing multi-LLM research strategies (Phase 1) and consolidating multi-model outputs into actionable intelligence (Phase 2).

1. Purpose

This skill provides 9 core capabilities:

| # | Capability | Phase | Description | |---|------------|-------|-------------| | 1 | Decompose | 1 | Break research questions into MECE structures | | 2 | Assign | 1 | Map question categories to optimal LLMs | | 3 | Assess | 1 | Evaluate research risks at appropriate depth | | 4 | Generate | 1 | Produce model-specific optimized prompts | | 5 | Consolidate | 2 | Synthesize multi-model outputs into unified findings | | 6 | Resolve | 2 | Handle conflicting information with WWHTBT protocol | | 7 | Classify | 2 | Score evidence quality and tag uncertainty types | | 8 | Detect | 2 | Identify coverage gaps and unknown unknowns | | 9 | Produce | 2 | Generate tiered, decision-ready research reports |

Checkpoints

This skill uses interactive checkpoints (see references/checkpoints.yaml) to resolve ambiguity:

  • research_type_classification — When research type is ambiguous
  • risk_depth_selection — When risk assessment depth not specified
  • model_mode_selection — When model execution mode not specified
  • hypothesis_priors_required — When multi_hypothesis enabled but priors missing
  • conflict_resolution_approach — When model outputs have significant conflicts (Phase 2)

2. Two-Phase Workflow

Phase 1: Research Design (Before Research)

| Step | Action | Output | |------|--------|--------| | 1 | Validate Objective | Confirm research question is answerable | | 2 | Classify Research Type | market | competitive | technology | strategic | | | CHECKPOINT: research_type_classification | If type ambiguous: AskUserQuestion | | 3 | Define Scope | In-scope, out-of-scope, boundaries | | 4 | Select MECE Pattern | 5-category decomposition structure | | 5 | Generate Sub-Questions | 3-4 questions per category | | 6 | Assess Risks | Quick | Standard | Comprehensive | | | CHECKPOINT: risk_depth_selection | If depth not specified: AskUserQuestion | | 7 | Assign Models | Map categories to Claude/Gemini/GPT | | | CHECKPOINT: model_mode_selection | If mode not specified: AskUserQuestion | | 8 | Frame Hypotheses | If multi_hypothesis=true | | | CHECKPOINT: hypothesis_priors_required | If priors missing: AskUserQuestion | | 9 | Recommend Expert Panel | If expert_panel=true | | 10 | Produce Research Brief | XML-structured Phase 1 deliverable |

Phase 2: Consolidation (After Research)

| Step | Action | Output | |------|--------|--------| | 1 | Ingest Model Outputs | Parse all LLM research results | | 2 | Score Evidence | Apply 5-point Evidence Strength Rubric | | 3 | Detect Conflicts | Identify where models disagree | | 4 | Resolve Conflicts | Apply WWHTBT for unresolved | | 5 | Classify Uncertainty | Tag as epistemic/aleatory/model | | 6 | Audit MECE Coverage | Check for coverage gaps | | 7 | Probe Unknown Unknowns | Run 5 discovery probes | | 8 | Tier Findings | Assign to Tier 1/2/3 by confidence | | 9 | Build Decision Support | Create if-then decision tree | | 10 | Define Kill Criteria | Conditions that invalidate research | | 11 | Produce Report | XML-structured Phase 2 deliverable |

3. Parameters

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | research_objective | string | required | The core research question or goal | | research_type | enum | market | market | competitive | technology | strategic | | model_mode | enum | parallel | parallel | sequential | convergent | | openai_depth | enum | balanced | minimal | balanced | exhaustive | | risk_depth | enum | standard | quick | standard | comprehensive | | multi_hypothesis | bool | false | Enable hypothesis-driven framing | | expert_panel | bool | false | Include expert panel recommendations | | context | string | "" | Additional context for research |

4. Model Strengths & Assignment

Model Profiles

| Model | Primary Strength | Best For | Limitation | |-------|------------------|----------|------------| | Claude Opus 4.5 | Judgment, synthesis, nuance | Strategic questions, conflict resolution, synthesis | May not surface all sources | | Gemini Pro 3 | Breadth, citations, grounding | Factual lookup, comprehensive sourcing, current data | Less depth on complex reasoning | | GPT-5.2 Deep | Recency, depth, exhaustiveness | Technical details, narrow deep-dives, edge cases | Can miss broader context |

Default Category Assignments

| Research Type | Claude | Gemini | GPT | |---------------|--------|--------|-----| | Market | Demand, Trends | Size, Structure, Supply | — | | Competitive | Positioning, Strategy | Product, GTM, Org | Deep Dive | | Technology | Fit, Risk | Maturity, Cost | Capability | | Strategic | Options, Stakeholders | Environment | Implementation |

5. Risk Assessment Depths

Quick (5 Factors)

Basic risk identification for time-sensitive research:

  • Top 3 risks with likelihood/impact
  • No mitigations or scenarios

Standard (+ Bias Audit)

Adds mitigation planning and cognitive bias check:

  • Mitigations and contingencies per risk
  • Early warning signals
  • Bias audit: confirmation, availability, anchoring

Comprehensive (+ Base Rates)

Full risk analysis with historical grounding:

  • Risk scenarios with trigger conditions
  • Risk dependencies and cascades
  • Base rate comparison from similar research
  • Pre-mortem analysis

6. MECE Decomposition Patterns

Pattern 1: Market Research

| Category | Focus | Model | |----------|-------|-------| | Market Size & Dynamics | TAM/SAM/SOM, growth rates | Gemini | | Market Structure | Segmentation, value chain | Gemini | | Demand Characteristics | Buyers, use cases, criteria | Claude | | Supply & Competition | Players, barriers, substitutes | Gemini | | Market Evolution | Trends, regulatory, disruption | Claude |

Pattern 2: Competitive Intelligence

| Category | Focus | Model | |----------|-------|-------| | Product & Offering | Features, pricing, roadmap | GPT | | Customers & Positioning | Segments, win/loss, messaging | Claude | | Go-to-Market | Sales, marketing, partnerships | Gemini | | Organization & Operations | Team, tech stack, cost structure | Gemini | | Strategy & Trajectory | Direction, investments, SWOT | Claude |

Pattern 3: Technology Evaluation

| Category | Focus | Model | |----------|-------|-------| | Capability & Performance | Features, benchmarks, limits | GPT | | Maturity & Ecosystem | Stability, community, tools | Gemini | | Fit & Integration | Use case alignment, migration | Claude | | Cost & Investment | TCO, licensing, infrastructure | Gemini | | Risk & Governance | Technical, vendor, compliance | Claude |

Pattern 4: Strategic Research

| Category | Focus | Model | |----------|-------|-------| | Current State | Position, strengths, weaknesses | Claude | | External Environment | Industry, macro, technology | Gemini | | Strategic Options | Directions, trade-offs, requirements | Claude | | Stakeholder Considerations | Customer, competitor, employee | Claude | | Implementation Requirements | Capabilities, investments, timeline | GPT |

7. Multi-Hypothesis Framing

When to Enable

  • Testing predictions or forecasts
  • Evaluating competing theories
  • Decision involves binary or multi-way choice
  • Need to avoid confirmation bias

Process

  1. Define core question as testable prediction
  2. Generate 2-4 MECE hypotheses covering all outcomes
  3. Assign prior probabilities (must sum to 100%)
  4. Define supporting and refuting evidence for each
  5. Research gathers evidence against criteria
  6. Update posteriors based on evidence strength

Example

<hypotheses question="Will enterprise adopt GenAI for customer service by 2027?">
  <hypothesis id="H1" position="broad" prior="30%">
    >50% enterprise adoption
  </hypothesis>
  <hypothesis id="H2" position="selective" prior="50%">
    10-50% adoption in specific use cases
  </hypothesis>
  <hypothesis id="H3" position="limited" prior="20%">
    <10% adoption due to barriers
  </hypothesis>
</hypotheses>

8. Evidence Strength Tribunal

5-point scale for evaluating source quality:

| Score | Name | Definition | Examples | |-------|------|------------|----------| | 5 | Primary | Direct from entity being researched | SEC filings, earnings calls, official docs | | 4 | Auth. Secondary | Major analysts with citations | Gartner, Forrester, WSJ investigative | | 3 | Credible Secondary | Reputable sources, some sourcing | TechCrunch, industry publications | | 2 | Weak Secondary | Unsourced, outdated, anonymous | LinkedIn self-reports, old reports | | 1 | Speculative | No verifiable basis | Rumors, predictions, fabrications |

Time Decay: Apply -1 for technology data >6 months, market data >1 year.

Reference: See references/evidence-strength-rubric.md for full scoring guidelines.

9. Conflict Resolution: WWHTBT

When models or sources disagree and resolution isn't clear, apply What Would Have To Be True analysis:

<conflict claim="Market size for X">
  <position holder="Gartner" value="$50B">
    <evidence score="4">2024 market report with methodology</evidence>
  </position>
  <position holder="IDC" value="$35B">
    <evidence score="4">Different scope definition</evidence>
  </position>

  <wwhtbt>
    <for_gartner>
      <condition>Adjacent markets included in scope</condition>
      <condition>Projected vs. realized revenue counted</condition>
    </for_gartner>
    <for_idc>
      <condition>Only core product category</condition>
      <condition>Realized revenue only</condition>
    </for_idc>
  </wwhtbt>

  <recommendation>
    Report range ($35-50B) with scope dependency noted.
    For our purposes, IDC definition more aligned.
  </recommendation>
</conflict>

10. Uncertainty Decomposition

| Type | Definition | Can Reduce? | Action | |------|------------|-------------|--------| | Epistemic | Knowledge gaps that COULD be closed | YES | Research further | | Aleatory | Inherent randomness that CANNOT be predicted | NO | Quantify range, build scenarios | | Model | Framework/definition dependencies | DEPENDS | Make choices explicit |

Classification Questions

  • Epistemic: "Does someone, somewhere know this?"
  • Aleatory: "Even with perfect info, would this still be uncertain?"
  • Model: "Would a different definition change the answer?"

Reference: See references/uncertainty-taxonomy.md for full classification protocol.

11. Gap Analysis

Part 1: MECE Coverage Audit

Compare findings against expected coverage matrix for research type. Flag:

  • Critical gaps: Core dimensions missing or Score ≤2
  • Significant gaps: Supporting dimensions weak
  • Minor gaps: Context items missing

Part 2: Unknown Unknowns Probes

| Probe | Question | |-------|----------| | Adjacent Domain | What lessons from related industries apply? | | Stakeholder Blind Spot | Whose voice is missing from sources? | | Time Horizon | What historical precedents or future implications are ignored? | | Failure Mode | What would have to be true for conclusions to be wrong? | | Second-Order Effects | If findings are true, what else must follow? |

Reference: See references/gap-analysis-protocol.md for full audit process.

12. Output Specifications

Phase 1 Deliverable: Research Brief

research-brief.xml
├── Header (ID, type, mode, parameters)
├── Section 1: Research Classification
├── Section 2: MECE Question Decomposition
├── Section 3: Multi-Hypothesis Framing (if enabled)
├── Section 4: Risk Assessment
├── Section 5: Expert Panel (if enabled)
├── Section 6: Model Role Assignments
├── Section 7: Ready-to-Execute Prompts
├── Section 8: Consolidation Strategy
├── Section 9: Verification Priorities
└── Section 10: Effort Estimates

Phase 2 Deliverable: Consolidated Report

consolidated-report.xml
├── Header (quality summary)
├── Part 1: Executive Summary (≤5 findings, bottom line)
├── Part 2: Tiered Findings (1: >75%, 2: 50-75%, 3: <50%)
├── Part 3: Evidence Quality Assessment
├── Part 4: Contested Claims & Conflict Resolution
├── Part 5: Uncertainty Analysis
├── Part 6: Gap Analysis
├── Part 7: Model Contribution Analysis
├── Part 8: Decision Support (if-then tree)
├── Part 9: Kill Criteria
├── Part 10: Methodology Transparency
├── Part 11: Appendices
└── CRITICAL CONSTRAINTS (at end for context retention)

Templates: See templates/research-brief-template.md and templates/consolidated-report-template.md

13. Expert Panel Integration

When to Enable

  • High-stakes decisions
  • Multi-disciplinary topics
  • Need for challenge/red-teaming
  • Regulatory or compliance implications

Process

  1. Identify panel size (3-8 experts) and balance
  2. Select domain-appropriate experts
  3. Define deliberation format (round-robin, debate, Delphi)
  4. Assign challenger role for assumption testing
  5. Synthesize panel perspectives into findings

Expert Selection by Domain

| Domain | Recommended Experts | |--------|---------------------| | Market | Market analyst, Customer representative, Industry veteran | | Competitive | Competitive intel analyst, Former competitor employee, Sales leader | | Technology | Technical architect, Security specialist, Operations lead | | Strategic | Strategy consultant, Board member, Industry analyst |

14. Quality Gates

Phase 1 Gates (Research Design)

| # | Gate | Criterion | |---|------|-----------| | 1 | Objective Clarity | Single, answerable research question | | 2 | MECE Validity | Categories non-overlapping and exhaustive | | 3 | Question Quality | All sub-questions researchable | | 4 | Model Fit | Assignments match model strengths | | 5 | Prompt Executability | Prompts can run without modification | | 6 | Completeness | All required sections populated |

Phase 2 Gates (Consolidation)

| # | Gate | Criterion | |---|------|-----------| | 1 | Evidence Scored | All findings have evidence scores | | 2 | Conflicts Surfaced | No hidden disagreements | | 3 | Uncertainty Classified | All gaps tagged by type | | 4 | Coverage Audited | MECE matrix reviewed | | 5 | Probes Executed | ≥3 of 5 unknown-unknowns probes run | | 6 | Tiers Justified | Confidence matches evidence profile | | 7 | Decision Support | Actionable if-then structure | | 8 | Constraints Verified | All 7 critical constraints checked |

15. Use Cases

| Use Case | Type | Mode | Risk | Hypothesis | Panel | |----------|------|------|------|------------|-------| | Market sizing | market | parallel | quick | no | no | | Competitor deep-dive | competitive | sequential | standard | no | no | | Build vs buy | technology | convergent | comprehensive | yes | yes | | Strategic planning | strategic | parallel | comprehensive | yes | yes | | Trend monitoring | market | parallel | quick | no | no | | Investment due diligence | competitive | convergent | comprehensive | yes | yes |

16. Workflow Integration

This skill integrates with the broader research workflow:

┌─────────────────────┐
│ research-interviewer│  Elicit research requirements
└──────────┬──────────┘
           │
           ▼
┌─────────────────────┐
│create-research-brief│  ◀── THIS SKILL (Phase 1)
│     (Phase 1)       │  Design multi-LLM research strategy
└──────────┬──────────┘
           │
           ▼
┌─────────────────────┐
│   Execute Research  │  Run prompts across models
│  (Manual or Agent)  │
└──────────┬──────────┘
           │
           ▼
┌─────────────────────┐
│create-research-brief│  ◀── THIS SKILL (Phase 2)
│     (Phase 2)       │  Consolidate into report
└──────────┬──────────┘
           │
           ▼
┌─────────────────────┐
│ consolidate-research│  Additional synthesis if needed
└─────────────────────┘

17. References and Templates

Reference Files

| File | Purpose | |------|---------| | references/evidence-strength-rubric.md | 5-point evidence scoring with special cases | | references/uncertainty-taxonomy.md | 3 uncertainty types with classification protocol | | references/gap-analysis-protocol.md | MECE audit + 5 unknown-unknowns probes | | references/mece-decomposition-guide.md | Full decomposition patterns with examples |

Template Files

| File | Purpose | |------|---------| | templates/research-brief-template.md | Phase 1 output structure (XML) | | templates/consolidated-report-template.md | Phase 2 output structure (XML) |

Quick Start

Phase 1: Create Research Brief

/create-research-brief
research_objective: "What is the market opportunity for AI legal research tools?"
research_type: market
risk_depth: standard

Phase 2: Consolidate Research

/create-research-brief --phase=2
input: [model outputs from Phase 1 execution]