baphuongna/pi-crew/council

Council Pattern — Adversarial Multi-Perspective Decision Making

When to Use

• Evaluating architecture decisions with significant tradeoffs
• Reviewing security-sensitive design choices
• Validating implementation plans before execution
• Resolving ambiguity where multiple valid approaches exist
• Deciding whether to build, buy, or extend

Prerequisites

• A clearly formulated question or decision to evaluate
• Sufficient context about the system for meaningful analysis

Operating Rules

1. You MUST spawn exactly 3 subagents with isolated context (fresh, not forked)
2. Each subagent receives ONLY the question — NO conversation history (anti-anchoring)
3. You MUST NOT influence any subagent's analysis direction
4. You MUST record all 3 votes before forming consensus
5. You MUST include dissent in the final recommendation

Workflow

Step 1: Formulate the Question

Write a clear, neutral question that includes:

• The decision to be made
• Relevant constraints (performance, security, timeline)
• Available options (if known)
• What "success" looks like

DO NOT bias the question toward any particular answer.

Step 2: Spawn 3 Council Members

Launch 3 parallel subagents, one per seat, using the councillor agents:

• Skeptic (finds flaws) → subagent_type='councillor-skeptic'
• Pragmatist (weighs tradeoffs) → subagent_type='councillor-pragmatist'
• Critic (stress-tests reasoning) → subagent_type='councillor-critic'

Each subagent's prompt is EXACTLY the question text from Step 1 — nothing else, no context preamble, no your-opinion-matters framing.

The seat perspective, output contract (Position/Confidence/Reasoning + seat-specific field), and anti-anchoring are baked into the agent files. The councillor agents carry inheritProjectContext: false, so isolation is STRUCTURAL (enforced at spawn), not merely instructed — do not wrap the question with your own framing or history.

All three seats return the shared vote block:

Position: PRO | CON | ABSTAIN
Confidence: 0.0-1.0
Reasoning: <analysis>
<Seat-specific field: Top 3 Risks | Alternatives Considered | Hidden Assumptions>

Step 3: Aggregate Votes

Collect all 3 responses. Compute consensus:

Step 4: Output Recommendation

## Council Decision: [Question Summary]
 
### Votes
| Role | Position | Confidence |
|------|----------|------------|
| Skeptic | PRO/CON/ABSTAIN | 0.X |
| Pragmatist | PRO/CON/ABSTAIN | 0.X |
| Critic | PRO/CON/ABSTAIN | 0.X |
 
### Consensus: [STRONG ACCEPT | WEAK ACCEPT | NO CONSENSUS | WEAK REJECT | STRONG REJECT]
 
### Recommendation
[One-paragraph synthesis]
 
### Key Insights
-[Best point from Skeptic]
-[Best point from Pragmatist]
-[Best point from Critic]
 
### Dissent
[Summary of any dissenting opinions and why they were overruled or remain unresolved]
 
### Action Items
-[ ] [Specific next step 1]
-[ ] [Specific next step 2]

Anti-Patterns

• DO NOT spawn fewer than 3 roles
• DO NOT share one subagent's analysis with another (contamination)
• DO NOT phrase the question to favor a specific outcome
• DO NOT override the council's consensus without documented justification
• DO NOT use council for trivial decisions (wastes resources)

Enforcement — Council Gate

Before finalizing a council result, verify:

• [ ] All 3 roles spawned with isolated (fresh) context
• [ ] Each role received ONLY the question, no prior conversation
• [ ] All 3 votes recorded with confidence scores
• [ ] Consensus level computed from vote pattern
• [ ] Dissent explicitly documented (not hidden)
• [ ] Recommendation includes actionable next steps

Budget

This skill applies a 3-attempt budget: 1 initial + max 2 re-attempts.

Stamp every invocation:

attempt X of 3 (Y attempts remaining)

An attempt is one full council round (3 adversarial agents → aggregate → consensus). Re-attempt when votes deadlock or dissent reveals information the first round missed.

Re-attempts only when the previous attempt materially changes the decision or risk. Do NOT spend a re-attempt on mechanical changes or already-resolved findings. When exhausted, escalate to the user with options (accept risk / change scope / exceptional budget).