← ClaudeAtlas

council-protocollisted

Use when several agents must deliberate on one question and their views must be combined into a single, honest synthesis with recorded dissent.
ymurat/academic-agent-skills · ★ 0 · AI & Automation · score 72
Install: claude install-skill ymurat/academic-agent-skills
# Council protocol Run after `panel-factory`. The moderator runs the steps and writes the synthesis; it does not vote. ## Evidence in one paragraph Most of the gain from multi-agent debate comes from independent answers plus majority vote, not from extra debate rounds (Debate or Vote, NeurIPS 2025, https://arxiv.org/abs/2508.17536). Debate methods do not reliably beat self-consistency and are sensitive to settings (https://arxiv.org/abs/2311.17371). Voting suits reasoning tasks, consensus suits knowledge tasks (ACL Findings 2025, https://arxiv.org/abs/2502.19130). Adversarial roles reduce early fixation (MAD, EMNLP 2024, https://arxiv.org/abs/2305.19118). Confidence-weighted rounds across diverse models help (ReConcile, https://arxiv.org/abs/2309.13007). So: protect independence, keep debate short, preserve dissent. ## Steps 1. Independent first round. Each member answers alone with: position, numbered findings (weaknesses first), evidence, open questions, confidence 1 to 5. No member sees another's answer. 2. Early stop check. If at least 4 of 5 members (80 percent) agree on the position and no member flags a factual error, skip to step 5. 3. Anonymized critique. Relabel answers A, B, C...; remove names and role hints. Each member ranks the others and lists the strongest and weakest point of each. 4. One rebuttal round at most, only on points of disagreement. A member may change position only by citing new evidence or a specific error; "the majority thinks so" is not a re