| name | council |
| description | Consult a multi LLM council for deliberation, debate, voting, critique, or verification. Uses GPT 5.4, Gemini 2.5, and Claude as peers. |
| user_invocable | true |
LLM Council Skill
You have access to the LLM Council MCP tools. Use them to orchestrate multimodel deliberation.
Available Tools
- council_deliberate. Full council deliberation with configurable protocol
- council_vote. Quick voting: all models answer, then anonymously rank each other
- council_debate. Structured debate with adaptive stopping
- council_critique. Peer critique or adversarial redteaming
- council_verify. Multi agent verification of an answer
- council_estimate_cost. Estimate cost before running
- council_status. Check which providers are available
- council_configure. Update default council composition
Usage Patterns
Quick consensus
Use council_vote for straightforward questions where you want the best answer selected by peer review.
Deep analysis
Use council_debate with adaptiveStop: true for complex questions that benefit from iterative refinement. Set maxRounds: 3 for thorough analysis.
Verification
Use council_verify to check if an answer is correct by having multiple models independently verify it.
Stress testing
Use council_critique with redTeam: true to find flaws, edge cases, and failure modes in responses.
Full deliberation
Use council_deliberate with protocol: "synthesize" for the chairman synthesis pattern. All models answer, then a chairman produces an authoritative synthesis.
Best Practices (from research)
- Scale agents (more models), not rounds (more debate turns)
- Heterogeneous models (mix providers) outperform homogeneous ones
- Anonymous peer review prevents model favoritism bias
- Always check cost estimate before large council runs
- Default council (GPT 5.4 + Gemini 2.5 Pro + Claude Sonnet 4.6) provides optimal diversity
Example Invocations
User: "What are the tradeoffs of microservices vs monolith?"
→ Use council_vote for quick consensus
User: "Is this algorithm correct? [code]"
→ Use council_verify
User: "Debate whether Rust or Go is better for this use case"
→ Use council_debate with maxRounds: 2
User: "Red team this API design"
→ Use council_critique with redTeam: true