| name | multi-perspective-deliberation |
| description | Coordinate a multi-persona deliberation session to challenge assumptions, debate architectural decisions, and perform pre-mortems. Leverages concurrent delegated agents/workers when the environment supports them, or falls back to single-context simulated dialogue. Trigger phrases: 'party mode', 'deliberation', 'multi-perspective', 'get a second opinion', 'debate this', 'challenge design'. |
Multi-Perspective Deliberation & Deliberative Dialogue
A framework to challenge designs, stress-test plans, and synthesize multi-disciplinary viewpoints (Product, Architecture, Development, Security, and Adversarial) before finalizing critical decisions or committing code.
Deliberation Personas
| Persona | Role | Key Lens |
|---|
| Product Owner | User & Value Advocate | "Does this solve a real user problem? Is the value worth the complexity?" |
| Architect | Design & Maintainability | "Is the architecture robust, decoupled, scalable, and elegant?" |
| Developer | Implementation Reality | "Can we build this easily? What are the hidden complexities and debt?" |
| Security Auditor | Trust & Security Skeptic | "What are the attack vectors, input hazards, and trust boundary risks?" |
| Devil's Advocate | Adversarial Challenger | "Why is this approach guaranteed to fail? What are the blind spots?" |
Phase 1: Execution Mode Selection
Determine the capabilities of the current execution environment:
Mode A: Concurrent Delegated Deliberation (Preferred)
Use when the environment supports parallel delegated agents/workers plus a way to exchange prompts, intermediate positions, or results between them.
- Launch Persona Workers: Create separate, concurrent delegated workers/agents for the selected personas (e.g., Architect, Developer, Security). Give each worker a clear persona brief and scope.
- Orchestrate Dialogue: Share the topic with all active workers. Rotate turns by relaying the other personas' positions, allowing them to respond to the topic and directly challenge each other's reasoning.
- Facilitate: Prompt the workers to address each other's concerns, surface trade-offs, and refine their positions.
Mode B: Simulated Single-Context Dialogue (Fallback)
Use in single-agent environments.
- Adopt the persona roles sequentially in turn-by-turn dialogue within a single context window.
- Ensure you represent each perspective with authentic disagreement. Avoid having the personas agree too quickly.
Phase 2: Deliberation Flow
- Setup: Identify the topic (or proposed implementation plan) and select 2-4 most relevant personas (default: Architect, Developer, Security).
- Opening Positions: Each persona state their initial stance on the topic (2-3 sentences max).
- Discussion Rounds:
- Personas respond to the topic and build or challenge other personas (e.g., "Building on what Architect said, but Developer's concern about timeline is valid because...").
- Ensure turns are concise (no monologues) and the conversation remains balanced.
- Introduce user interjections if running in human-in-the-loop mode.
- Convergence: Each persona delivers their final adjusted position, noting what they agree on, what they still contest, and what they need to investigate.
- Synthesis: Produce a consensus and divergence summary.
Phase 3: Deliberation Synthesis Output
Compile the deliberation results into the following structure:
# Deliberation Report: [Topic]
- **Deliberation Mode**: [Delegated Multi-Agent / Simulated Dialogue]
- **Participants**: [List of active personas]
## 1. Consensus (Points of Agreement)
- [List specific designs, boundaries, or constraints everyone agreed on]
## 2. Divergence (Unresolved Tensions)
- **[Persona A] vs. [Persona B]**: Describe the core conflict (e.g. security overhead vs. development speed).
## 3. Pre-Mortem (Failure Scenarios)
*If this plan fails in 6 months, why did it fail?*
- [Scenario A with mitigation]
## 4. Final Verdict & Action Items
- **Verdict**: [PROCEED / REVISE / RETHINK]
- **Action items**: [Specific tasks to resolve key disputes]
