Expert Panel Deliberation

Structured multi-expert evaluation, deliberation, and consensus building.

Purpose

Provide consistent, high-quality multi-perspective analysis by:

• Instantiating domain-appropriate expert panels
• Executing structured deliberation protocols
• Resolving conflicts between viewpoints
• Building weighted consensus
• Documenting reasoning and dissent

When to Use

Ideal for:

• Decisions requiring multiple perspectives
• Evaluating options with tradeoffs
• Complex topics with no single right answer
• Building confidence through diverse viewpoints

Avoid when:

• Simple factual questions
• Time-critical decisions needing speed
• Single-domain technical questions

Checkpoints

This skill uses interactive checkpoints (see

references/checkpoints.yaml

• domain_context — When domain not specified
• output_format_selection — When output format not specified
• deliberation_depth_selection — When depth could vary based on stakes
• panel_composition — When archetype selection is ambiguous
• panel_size_adjustment — When specified size conflicts with recommended
• conflict_resolution_approach — When unresolved conflicts need user input
• weighting_approach — When consensus weighting is ambiguous
• dissent_documentation — When minority view significance is unclear

Workflow

Step 1: Define Panel Requirements

Determine what evaluation is needed:

• Subject: What is being evaluated?
• Goal: What should the evaluation determine?
• Panel size: 3-8 experts (default: 5)
• Output format: findings | scores | ranking | recommendation

CHECKPOINT: domain_context

• If domain not specified and not inferable from subject: AskUserQuestion
• Options: Architecture, Product, Security, Operations, Business, Custom
• Example: "What domain is this evaluation for?"

CHECKPOINT: output_format_selection

• If output format not specified: AskUserQuestion
• Options: Findings, Scores, Ranking, Recommendation
• Example: "What output format would be most useful?"

CHECKPOINT: deliberation_depth_selection

• If stakes unclear and depth not specified: AskUserQuestion
• Options: Quick, Standard, Deep
• Example: "How thorough should the deliberation be?"

Step 2: Assemble Expert Panel

Select from archetypes based on domain:

Panel composition rules:

• Minimum 3 experts for meaningful deliberation
• Maximum 8 experts (diminishing returns beyond)
• Always include at least one challenger perspective
• Balance technical and business viewpoints

CHECKPOINT: panel_composition

• If subject matches multiple domains or custom panel needed: AskUserQuestion
• Options: Architecture panel, Product panel, Security panel, Custom selection
• Example: "Which expert panel composition fits best?"

CHECKPOINT: panel_size_adjustment

• If user-specified size differs significantly from recommended: AskUserQuestion
• Example: "You specified 3 experts, but this topic typically benefits from 5. Which size?"

Step 3: Execute Individual Evaluation

For each expert:

1. Adopt perspective — Review role, expertise, concerns
2. Evaluate subject — Answer from expert's viewpoint
3. Score (if applicable) — Rate relevant dimensions
4. Document — Key findings, concerns, recommendations

Step 4: Execute Deliberation

Round 1: Finding Presentation

• Each expert presents top findings
• No debate yet; just surface perspectives

Round 2: Cross-Examination

• Experts question each other's findings
• Surface disagreements and gaps

Round 3: Conflict Resolution

• Address disagreements systematically
• Document resolved vs. unresolved conflicts

Round 4: Consensus Building

• Identify areas of agreement
• Weight by expert influence
• Synthesize combined view

CHECKPOINT: conflict_resolution_approach

• If significant conflicts remain unresolved: AskUserQuestion
• Options: Weight by domain, Weight by evidence, Document both, Escalate to user
• Example: "Technical Authority and User Advocate disagree. How should we resolve?"

CHECKPOINT: weighting_approach

• If weighting algorithm not specified and experts have unequal relevance: AskUserQuestion
• Options: Equal weight, Domain relevance, Confidence-adjusted, Evidence-based
• Example: "How should expert perspectives be weighted?"

Step 5: Generate Output

Format based on requested output type.

CHECKPOINT: dissent_documentation

• If minority views exist and significance unclear: AskUserQuestion
• Options: Document prominently, Document briefly, Omit
• Example: "Risk Specialist disagrees with consensus. How should we document this?"

Output Format

## Expert Panel Analysis: [Subject]

### Panel Composition
- **[Expert 1 Role]:** [brief expertise]
- **[Expert 2 Role]:** [brief expertise]
...

### Key Findings

**Consensus Views:**
1. [Finding agreed by most/all experts] (Confidence: HIGH/MED/LOW)
2. [Finding agreed by most/all experts] (Confidence: HIGH/MED/LOW)

**Divergent Views:**
- [Expert] believes [X] while [Expert] believes [Y]
  - Resolution: [how addressed or "unresolved"]

### Scores (if applicable)

| Dimension | Score | Confidence |
|-----------|-------|------------|
| [dim] | [X/10] | [H/M/L] |

### Recommendations

1. [Primary recommendation]
2. [Secondary recommendation]

### Dissent Record

[Any unresolved disagreements and minority views]

Quality Gates

• All requested experts contributed
• Each expert answered from their perspective
• Conflicts identified and addressed
• Weights properly applied
• Consensus clearly stated
• Dissent documented if present
• Confidence levels assigned
• All applicable checkpoints evaluated (ambiguity resolved via AskUserQuestion)

Parameters

panel_size

deliberation_depth

output_format

include_challenger

Examples

Example 1: Architecture Decision User: Analyze our microservices vs monolith decision with an expert panel Panel: Technical Authority, Quality Guardian, Risk Specialist, Efficiency Expert, Challenger [Deliberation proceeds with each expert evaluating from their perspective]

Example 2: Feature Prioritization User: Get expert perspectives on which features to build next quarter Panel: User Advocate, Domain Specialist, Technical Authority, Efficiency Expert, Challenger [Deliberation with scoring output]