Council Skill

Runs a panel of N agents with opposing perspectives on a shared artifact (specs, design docs, architecture decisions). Each agent has a persona, a mandate, and permission to block. Returns a synthesized verdict.

When to use

• Reviewing specs before dispatching implementation agents
• Evaluating architecture decisions with genuine tradeoffs
• Stress-testing a TRIZ resolution from multiple lenses
• Any decision where a single perspective is insufficient

When NOT to use

• Code review (use pr-review-toolkit)
• Merge conflicts (semantic problem, not adversarial)
• Simple factual questions (overkill)

Invocation

/council <artifact_path_or_description> [--n 3] [--personas <list>]

Default N=3. Default personas: Purist, Pragmatist, Adversary.

Default personas

Custom personas

Override with --personas for domain-specific councils:

/council specs/ --personas "Rust systems engineer, UX researcher, Security auditor"

Output format

Each agent returns:

VERDICT: APPROVE | APPROVE WITH CONDITIONS | BLOCK
BLOCKER: <one-line reason if blocking>
CONDITIONS: <list of required changes if conditional>
OBSERVATIONS: <key findings>
VOTE: <principle or approach they'd recommend>

Synthesizer agent reads all N verdicts and produces:

COUNCIL VERDICT: APPROVED | APPROVED WITH CONDITIONS | BLOCKED
BLOCKING AGENTS: <list>
REQUIRED CHANGES: <merged deduplicated list>
DISSENTS: <disagreements that didn't block>
SYNTHESIS: <what the council agrees on>

How to implement (for Claude Code)

1. Read the artifact(s)
2. Spawn N agents in parallel via Agent tool, each with persona prompt + artifact content
3. Collect all N verdicts
4. Run synthesizer pass on combined verdicts
5. Present council verdict to user

Persona prompt template

You are [PERSONA_NAME], a council member reviewing the following artifact.

Your mandate: [MANDATE]

You are adversarial. Your job is not to rubber-stamp — it is to find problems,
challenge assumptions, and block if necessary.

VERDICT options:
- APPROVE: No significant issues.
- APPROVE WITH CONDITIONS: Issues exist but are fixable. List CONDITIONS.
- BLOCK: Critical issue. State BLOCKER clearly in one sentence.

Artifact:
---
[ARTIFACT_CONTENT]
---

Respond in this exact format:
VERDICT: <APPROVE|APPROVE WITH CONDITIONS|BLOCK>
BLOCKER: <one sentence, or "none">
CONDITIONS: <bullet list, or "none">
OBSERVATIONS: <your key findings, 3-5 bullets>
VOTE: <what you'd recommend instead or in addition>

Notes

• Agents run in parallel (independent, not sequential deliberation)
• Synthesis is a separate pass — do not ask agents to read each other's verdicts
• A single BLOCK from any agent surfaces to the user as a council BLOCK
• The council is advisory — human has final say
• Log council verdicts to corpus when reviewing TRIZ resolutions: triz log --council-verdict council-output.json

Verification contract (auto-injected by proveryay)

Every agent dispatched through this skill receives the following appended to its prompt automatically:

<see ~/.claude/skills/proveryay/CLAIMS-TEMPLATE.md — injected verbatim>

This means every agent MUST emit a <claims> block when it makes any deployment/state-change assertion. The proveryay SubagentStop hook will verify each claim in parallel and inject structured verdicts into the dispatcher's next-turn context.

If the agent's work is purely planning / research / narrative (no state changes — typical for council verdicts), the agent may omit the <claims> block. Proveryay treats absence as silent (no noise).

Invocation example with template injection:

# Prompt injection: template auto-appended (idempotent)
PROMPT="${ORIGINAL_PROMPT}

$(cat ~/.claude/skills/proveryay/CLAIMS-TEMPLATE.md)"

See:

• ~/.claude/skills/proveryay/SKILL.md — hook behavior
• ~/.claude/skills/proveryay/CLAIMS-TEMPLATE.md — canonical format
• ~/.claude/skills/fleet-eval/SKILL.md — probe taxonomy