Back to Details

debate

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Debate — 多 AI 分身 P2P 辩论

Debate — Multi-AI Agent P2P Debate

派出 3 个 AI 分身（架构师、挑刺者、实战工程师）围绕一个议题进行 P2P 互发辩论，自动归档全部消息到文档。
Deploy 3 AI Agents (Architect, Skeptic, Practical Engineer) to conduct P2P mutual debate around a topic, and automatically archive all messages to a document.

前置检查

Pre-checks

开始前，确认 
~/.claude/settings.json
的 
env
中包含：
json
"CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
如果缺失，先让用户加上并重启 Claude Code。
Before starting, confirm that the 
env
in 
~/.claude/settings.json
includes:
json
"CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
If missing, ask the user to add it and restart Claude Code first.

流程

Process

1. 明确议题

1. Clarify the Topic

向用户确认辩论主题。一次问一个问题，不要轰炸。了解的维度：
  • 辩论的具体话题是什么？
  • 有没有特别关注的争议点？
  • 期望的辩论模式（见下方）？
Confirm the debate topic with the user. Ask one question at a time, don't bombard. Understand the following dimensions:
  • What is the specific debate topic?
  • Are there any particular controversial points to focus on?
  • Desired debate mode (see below)?

2. 选择模式

2. Select Mode

提供三种模式供用户选择：
模式角色适用场景
三视角（默认）架构师 + 挑刺者 + 实战工程师技术方案探索、架构设计讨论
对抗辩论正方 + 反方 + 裁判有明确二元选择时（选 A 还是 B）
魔鬼代言人提案者 + 挑刺者已有方案想被压力测试
Provide three modes for the user to choose from:
ModeRolesApplicable Scenarios
Tri-perspective (Default)Architect + Skeptic + Practical EngineerTechnical solution exploration, architecture design discussion
Adversarial DebateProponent + Opponent + RefereeWhen there is a clear binary choice (choose A or B)
Devil's AdvocateProposer + SkepticWhen an existing solution needs stress testing

3. 创建团队与分身

3. Create Team & Agents

TeamCreate → Agent × N（run_in_background: true）
每个分身的 prompt 必须包含：
  • 角色定义与讨论主题
  • 队友名称列表
  • 明确指令：写完分析后用 SendMessage 同时发给队友和 team-lead（抄送以记录全文）
  • 响应指令：收到队友消息后，直接回复他，并抄送 team-lead
  • 不准等待 team-lead 指令，自己主动发
TeamCreate → Agent × N（run_in_background: true）
The prompt for each agent must include:
  • Role definition and discussion topic
  • List of teammate names
  • Clear instruction: After completing the analysis, use SendMessage to send to both teammates and team-lead (CC to record the full text)
  • Response instruction: After receiving a message from a teammate, reply directly to them and CC the team-lead
  • Must not wait for team-lead instructions, take the initiative to send messages

4. 监控与记录

4. Monitoring & Recording

  • 所有 CC 给 team-lead 的消息，实时追加到项目根目录的 
    debate-<topic>.md
  • P2P 未抄送的消息只能看到摘要，无法记录全文——提醒分身抄送
  • 如果分身 3 次 idle 仍不产出内容，主动 SendMessage 催促
  • All messages CC'd to the team-lead are appended in real-time to 
    debate-<topic>.md
    in the project root directory
  • Only summaries of P2P messages without CC can be seen, full text cannot be recorded — remind agents to CC
  • If an agent remains idle 3 times without producing content, actively send a SendMessage to urge them

5. 辩论推进

5. Advance the Debate

  • 首轮：各分身独立发表初始观点
  • 后续：分身自发互评辩论。如果你发现某条线断了，手动 SendMessage 重新触发
  • 当辩论进入边际递减（重复论述、无新论点）时，提示用户收束
  • First round: Each agent independently presents initial views
  • Subsequent rounds: Agents spontaneously comment and debate. If you find a thread is broken, manually send a SendMessage to re-trigger
  • When the debate reaches diminishing returns (repeated arguments, no new points), prompt the user to conclude

6. 收束与归档

6. Conclude & Archive

  • 让每个分身给出最终结论
  • 整理共识框架和未决问题
  • 关闭分身（shutdown_request → TeamDelete）
  • 文档最终落盘
  • Ask each agent to provide a final conclusion
  • Organize the consensus framework and unresolved issues
  • Shut down agents (shutdown_request → TeamDelete)
  • Finalize the document

文档格式

Document Format

辩论文档应包含：
  • 前置配置说明
  • 每一轮的分身发言（标注发送方 → 接收方）
  • 辩论连通图
  • 最终共识框架
  • 关键结论表（争论点 / 收敛方向 / 状态）
  • 辩论统计数据
The debate document should include:
  • Pre-configuration instructions
  • Agent speeches for each round (marked with Sender → Receiver)
  • Debate connection graph
  • Final consensus framework
  • Key conclusion table (Controversial Point / Convergence Direction / Status)
  • Debate statistics

已知问题

Known Issues

  • 实验性功能，分身响应偶有不稳定
  • 分身可能不响应 shutdown，等 session 结束自动回收
  • 未 CC 的 P2P 消息无法获取全文
  • 分身不会主动发起多轮对话，需要 prompt 中明确指令
  • Experimental feature, agent responses are occasionally unstable
  • Agents may not respond to shutdown, will be automatically recycled when the session ends
  • Full text of P2P messages without CC cannot be obtained
  • Agents will not actively initiate multi-round conversations, clear instructions in the prompt are required