---
name: expert-panel
description: Structured multi-perspective review with dynamic panel selection. 12-seat roster, selection panel picks 3-7, up to 3 rounds, concrete verdict. Includes decision router, light mode, and second-order thinking (Devil's Advocate traces consequences forward).
version: 2.5.1
tools: []
requires:
  mcp: []
  optional_mcp: []
  memory: []
limitations: "All perspectives are generated by a single model. The panel surfaces considerations across predefined lenses but does not replicate independent expert judgment. Use peer review for irreversible decisions."
verdict_taxonomy: "3-tier (v2.5): Fix ranked, Defer with trigger, Out of scope"
triggers:
  - "expert panel"
  - "roundtable"
  - "red team this"
  - "blind spot check"
  - "stress test"
  - "critique this"
  - "what am I missing"
  - "poke holes"
  - "pre-mortem"
author: gesked
codeowner: gesked
origin: Evolved from fxschoen's ab-test edge-case reviewer. Expanded to full panel by gesked (v1.0). Decision router added v2.0. Second-order thinking (Devil's Advocate) added v2.1. Light mode added v2.2. Verdict taxonomy v2.3 introduced 4-tier. v2.4 collapsed to 3-tier (must/should/nice). v2.5 collapsed to 2-decision (Fix ranked + Defer) + Out of scope. v2.5.1 accessibility compliance (emoji removed), sentence case, LP Tension section restored, Reframe constrained, Responsible AI footer, AI writing cleaned, golden rubric format.
---

# Expert Panel v2.5.1

Structured multi-perspective review. A selection panel (3 seats) first decides who the right reviewers are, then a working panel (up to 7 seats from a 12-seat roster) evaluates your decision, document, or plan.

Up to 3 rounds. Each round surfaces new blind spots before locking a verdict.

Includes:
- Decision router (full panel vs light mode vs out of scope)
- Second-order thinking via Devil's Advocate lens
- LP Tension section: forces naming the active Leadership Principle tradeoff
- Responsible AI footer: single-model limitation disclosure
- Golden rubric format (7/7)

## Roster

12 named seats. Selection panel picks 3-7 per run.

1. Skeptic - Challenges assumptions and flags overconfidence
2. Builder - Evaluates feasibility, cost, and implementation risk
3. Customer Advocate - Centers user/customer impact and experience
4. Security Reviewer - Flags threat surface, trust boundaries, data exposure
5. Legal/Compliance - Identifies regulatory and contractual risk
6. Data Scientist - Questions statistical validity and measurement design
7. Product Strategist - Evaluates market fit, roadmap coherence, competitive risk
8. Operator - Focuses on runbooks, on-call burden, and operational readiness
9. Finance Reviewer - Validates cost models, ROI assumptions, and budget risk
10. Devil's Advocate - Traces second-order consequences and steelmans the opposition
11. Domain Expert - Provides deep subject-matter knowledge relevant to the context
12. Leadership Advisor - Applies Amazon Leadership Principles and org dynamics

## Decision Router

Before running the full panel, the router evaluates:
- Is this a decision, document, plan, or code artifact?
- Is the stakes level high enough to warrant 3-7 reviewers?
- Is the scope clear enough to run?

Routes to:
- Full panel (3-7 reviewers, up to 3 rounds)
- Light mode (2 reviewers, 1 round, for low-stakes or exploratory)
- Out of scope (factual lookup, personal opinion, not a reviewable artifact)

## Verdict Taxonomy (v2.5)

Three tiers - all Fix items are mandatory, no optional "shoulds":

**Fix** (ranked 1, 2, 3...): Must address before proceeding. Concrete, actionable, specific. Ranked by severity.

**Defer** (with trigger): Valid concern, not blocking. Includes a named trigger condition: "Revisit when X." No open-ended defers.

**Out of scope**: Outside the panel's mandate for this artifact.

## LP Tension Section

Every verdict names the active Leadership Principle tradeoff. Examples:
- "Bias for Action vs. Are Right A Lot"
- "Customer Obsession vs. Frugality"
- "Invent and Simplify vs. Earn Trust"

Forces explicit acknowledgment of the tradeoff being made, not just the decision.

## Reframe Rule

Each panelist may open with a premise challenge (reframe) before the verdict. Constrained to 2-3 sentences. Must come first. Cannot bloat into a full argument.

## Accessibility Compliance

- Zero emoji in output (WCAG 1.4.1, color-independent design)
- Sentence case throughout (AWS Style Guide rule 44080)
- Text labels carry full meaning without visual indicators

## Responsible AI Footer

All panel outputs include: "Note: All perspectives generated by a single model. This panel surfaces considerations across predefined lenses but does not replicate independent expert judgment. For irreversible decisions, supplement with actual peer review."

## Example Triggers

- "Expert panel on this doc"
- "Red team this plan"
- "Blind spot check before I send"
- "Pre-mortem on this project"
- "What am I missing here"
- "Roundtable on this decision"
- "Stress test my architecture"
- "Poke holes in this proposal"