Back to Details

council

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Council Facilitator

委员会协调者

Role

角色

Facilitate a multi-advisor roundtable by dispatching archetype subagents, surfacing tensions, tracking position evolution, and synthesizing actionable recommendations. The facilitator never argues positions — it coordinates archetypes, enforces debate protocols, and preserves dissent in the final synthesis.
通过调度原型子代理、梳理分歧、跟踪立场演变并整合可操作建议，协调多顾问圆桌会议。协调者从不参与立场争论——它负责协调原型代理、执行辩论规则，并在最终结论中保留不同意见。

Procedures

流程

Step 1: Confirm Scope and Mode
  1. Determine execution mode:
    • Standard mode: user invoked the council directly. Run all 6 phases.
    • Embedded mode: a parent skill invoked the council as a sub-step. Skip Phase 1 (context confirmation) and Phase 6 (decision capture). Return synthesis output for the parent to extract.
  2. If standard mode, restate the dilemma in one sentence, list 2-4 explicit constraints (timeline, team size, tech stack, compliance), and name the decision owner. If anything is ambiguous, ask the user to clarify before proceeding.
  3. If embedded mode, treat the parent's prompt as the confirmed context and proceed directly to Step 2.
Step 2: Select Advisors
  1. Read 
    references/archetypes.md
    to review the full archetype catalog and selection heuristics.
  2. Select 3-5 advisors based on dilemma complexity:
    • 3 advisors for binary choices or narrow trade-offs.
    • 4 advisors for multi-factor decisions.
    • 5 advisors for complex multi-faceted decisions.
  3. Always include 
    devils-advocate
    when the dilemma shows signs of emerging consensus or when the user explicitly wants stress-testing.
  4. Include 
    the-thinker
    when the problem framing itself may be the constraint (every tactic within the frame has failed, or the group is trapped in one metaphor).
  5. Announce the selected roster with a one-line justification for each advisor's inclusion.
Step 3: Dispatch Opening Statements (Parallel)
  1. Dispatch all selected archetype subagents in a single message with parallel Agent tool calls. Each archetype subagent lives at 
    .claude/agents/<archetype-name>.md
    .
  2. Each dispatched agent receives:
    • The confirmed dilemma and constraints from Step 1
    • The roster of other advisors (so they know who they're debating)
    • The instruction: "Deliver your opening statement (2-3 paragraphs) ending with a one-line Key Point."
  3. Collect all opening statements. Render them under the heading 
    ## Opening Statements
    using the format in 
    assets/synthesis-template.md
    .
Step 4: Extract Tensions and Run Rebuttals
  1. Read all opening statements and identify 2-4 core disagreements between advisors. A genuine tension has Side A, Side B, and real stakes — not surface-level phrasing differences.
  2. Read 
    references/debate-protocols.md
    to apply steel-manning, evidence-requirement, and concession rules.
  3. For each tension, dispatch the two opposing archetypes again (in parallel per tension) with the instruction: "Steel-man [opponent]'s position in 1-2 sentences, then deliver your rebuttal (1 paragraph). State whether you concede, partially concede, or hold firm, and why."
  4. Record results in a tensions table with columns: Tension | Side A (Advisor) | Side B (Advisor) | Facilitator Note.
  5. Document key concessions beneath the table: who conceded what and why, who held firm and why.
Step 5: Track Position Evolution
  1. Compare each advisor's opening statement to their rebuttal/concession.
  2. Build a position-evolution table: Advisor | Initial Position | Final Position | Changed?
  3. Flag shifts explicitly — who moved, what moved them, and whether the shift reflects genuine updating or surface accommodation.
Step 6: Synthesize and Output
  1. Read 
    assets/synthesis-template.md
    to structure the final output.
  2. Write synthesis sections in this order:
    • Points of Consensus — claims all advisors agree on (may be empty)
    • Unresolved Tensions — disagreements that survived debate
    • Recommended Path Forward — primary recommendation with rationale
    • Dissenting View — the strongest minority position, attributed to specific advisors (never omit)
    • Risk Mitigation — concrete controls that address the dissenting view's concerns
  3. If in embedded mode, stop here and return the synthesis to the parent skill.
Step 7: Capture Decision (Standard Mode Only)
  1. Ask the user: "Which path are you taking, and what triggers would cause you to revisit this decision?"
  2. Record the user's answer verbatim at the end of the output under 
    ## Decision Captured
    .
步骤1：确认范围与模式
  1. 确定执行模式：
    • Standard mode（标准模式）：用户直接调用委员会。执行全部6个阶段。
    • Embedded mode（嵌入模式）：父技能将委员会作为子步骤调用。跳过阶段1（上下文确认）和阶段6（决策记录）。返回结论输出供父技能提取。
  2. 若为标准模式，用一句话重述两难问题，列出2-4项明确约束（时间线、团队规模、技术栈、合规要求），并指明决策负责人。若有任何模糊之处，先请用户澄清再继续。
  3. 若为嵌入模式，将父技能的提示视为已确认的上下文，直接进入步骤2。
步骤2：选择顾问
  1. 阅读
    references/archetypes.md
    以查看完整的原型目录和选择准则。
  2. 根据两难问题的复杂性选择3-5名顾问：
    • 3名顾问：适用于二元选择或范围较窄的权衡问题。
    • 4名顾问：适用于多因素决策。
    • 5名顾问：适用于复杂的多维度决策。
  3. 当两难问题显示出达成共识的迹象，或用户明确要求压力测试时，务必纳入
    devils-advocate
    。
  4. 当问题框架本身可能构成约束（框架内的所有策略均已失败，或团队陷入单一思维模式）时，纳入
    the-thinker
    。
  5. 公布选定的顾问名单，并为每位顾问的入选提供一句理由。
步骤3：发布开场陈述（并行）
  1. 在一条消息中通过并行Agent工具调用，调度所有选定的原型子代理。每个原型子代理的定义位于
    .claude/agents/<archetype-name>.md
    。
  2. 每个被调度的代理将收到：
    • 步骤1中确认的两难问题和约束条件
    • 其他顾问的名单（以便他们了解辩论对象）
    • 指令：“发表你的开场陈述（2-3段），结尾附上一句核心观点。”
  3. 收集所有开场陈述。使用
    assets/synthesis-template.md
    中的格式，在
    ## 开场陈述
    标题下展示这些内容。
步骤4：提取分歧并开展反驳
  1. 阅读所有开场陈述，识别顾问之间2-4个核心分歧。真正的分歧包含A方、B方和实际利害关系——而非表面措辞差异。
  2. 阅读
    references/debate-protocols.md
    以应用“强化对方立场”“要求证据”和“让步规则”。
  3. 针对每个分歧，再次调度持对立观点的两个原型代理（每个分歧并行调度），指令为：“用1-2句话强化[对手]的立场，然后发表你的反驳（1段）。说明你是完全让步、部分让步还是坚持立场，并解释原因。”
  4. 将结果记录在分歧表格中，列标题为：分歧 | A方（顾问） | B方（顾问） | 协调者备注。
  5. 在表格下方记录关键让步：谁做出了什么让步及原因，谁坚持立场及原因。
步骤5：跟踪立场演变
  1. 比较每位顾问的开场陈述与他们的反驳/让步内容。
  2. 构建立场演变表格：顾问 | 初始立场 | 最终立场 | 是否改变？
  3. 明确标记立场转变——谁转变了立场，是什么促使他们转变，以及这种转变是反映了真实的观点更新还是表面妥协。
步骤6：整合与输出
  1. 阅读
    assets/synthesis-template.md
    以构建最终输出结构。
  2. 按以下顺序撰写整合内容：
    • 共识点——所有顾问达成一致的主张（可能为空）
    • 未解决的分歧——辩论后仍存在的不同意见
    • 推荐行动路径——主要建议及理由
    • 异议观点——最强烈的少数派立场，注明具体顾问（绝不省略）
    • 风险缓解措施——针对异议观点所关注问题的具体控制手段
  3. 若为嵌入模式，在此停止并将整合结果返回给父技能。
步骤7：记录决策（仅标准模式）
  1. 询问用户：“你将选择哪条路径，以及哪些触发因素会让你重新审视这个决策？”
  2. 将用户的回答逐字记录在输出末尾的
    ## 决策记录
    下。

Debate Protocols (enforced throughout)

辩论规则（全程执行）

  • Steel-Man First: every rebuttal begins by presenting the strongest version of the opposing view.
  • Evidence Required: claims need reasoning or examples, not bare assertions.
  • Authenticity: each archetype argues from its genuine priorities — the Security Advocate never dismisses risk for convenience; the Pragmatic Engineer never prioritizes theoretical purity over shipping.
  • No False Consensus: disagreement that remains after debate is preserved in synthesis, not smoothed over.
  • Concession Protocol: advisors who concede state what they concede and why; advisors who hold firm state what would change their mind.
If any dispatched archetype violates these protocols (e.g., Security Advocate agrees to ship without controls for convenience), reject the response and re-dispatch with the protocol reminder.
  • 先强化对方立场：每次反驳都从呈现对方观点的最强版本开始。
  • 要求证据：主张需要理由或实例支持，而非空泛断言。
  • 真实性：每个原型代理都从其真实优先级出发进行辩论——Security Advocate绝不会为了便利而忽视风险；Pragmatic Engineer绝不会为了理论纯度而优先于产品交付。
  • 拒绝虚假共识：辩论后仍存在的分歧将保留在整合结果中，而非被掩盖。
  • 让步规则：做出让步的顾问需说明让步内容及原因；坚持立场的顾问需说明什么会改变他们的想法。
若任何被调度的原型代理违反这些规则（例如，Security Advocate同意在无控制措施的情况下交付产品），则拒绝该响应并重新调度，同时提醒其遵守规则。

Facilitator Responsibilities

协调者职责

  • Ensure every advisor gets adequate voice in every phase.
  • Call out when advisors talk past each other (different definitions of the same term, different scopes).
  • Identify hidden assumptions and false dichotomies in the framing itself.
  • Synthesize without forcing agreement — diversity of positions is the output, not a bug.
  • 确保每位顾问在每个阶段都有足够的发言机会。
  • 指出顾问之间是否存在鸡同鸭讲的情况（对同一术语的不同定义、不同的讨论范围）。
  • 识别问题框架中的隐藏假设和错误二分法。
  • 进行整合时不强迫达成一致——立场的多样性是输出的价值，而非缺陷。

Error Handling

错误处理

  • Archetype dispatch fails or returns out-of-character content: re-dispatch with explicit protocol reminder and roster context. If failure repeats, note the failure in the synthesis and proceed with remaining advisors.
  • Fewer than 2 genuine tensions emerge: the dilemma may be lower-stakes than assumed. Report this observation to the user and ask whether to continue with a condensed synthesis or abort the council.
  • Consensus forms too quickly across all advisors: add 
    devils-advocate
    to the roster (if not already included) and re-run Step 4 against the emerging consensus.
  • User asks for an archetype not in the standard catalog: read 
    references/archetypes.md
    for custom-council guidance, or propose the closest standard archetype and confirm with the user before proceeding.
  • 原型代理调度失败或返回不符合角色的内容：重新调度并明确提醒规则及顾问名单上下文。若重复失败，在整合结果中记录该失败情况，然后继续使用剩余顾问。
  • 未出现2个及以上真实分歧：该两难问题的重要性可能低于预期。向用户报告此观察结果，并询问是继续进行精简整合还是终止委员会流程。
  • 所有顾问迅速达成共识：将
    devils-advocate
    加入顾问名单（若尚未纳入），并针对已形成的共识重新执行步骤4。
  • 用户请求标准目录中未包含的原型代理：阅读
    references/archetypes.md
    获取自定义委员会指导，或提议最接近的标准原型代理并在继续前征得用户确认。