模型委员会
MVPavan2026-10-27查看 921复制 1393
仅用于明确请求由指定模型或智能体组成评审团并设有裁判的情形。普通审查和头脑风暴无需评审团。
📘如何使用这个技能
- 1复制技能内容·点击上方「Copy SKILL.md」复制全部内容,或 Download 下载到本地
- 2打开你的 AI 客户端·支持 Claude Code、Trae、Cursor、ChatGPT、DeepSeek、Kimi 等任意 AI 助手
- 3粘贴为系统提示/知识文件·Claude Code:放入 .claude/skills/ 目录;普通对话:作为首条消息发送并说「请严格按这份技能文档执行」
- 4直接提出你的需求·AI 会按技能定义的工作流、格式和约束来处理,无需再解释背景
💡 提示:技能生效后,每次对话只需一句需求,AI 会按 SKILL.md 的流程执行。
SKILL.md2026 chars
--- name: model-council description: Use only for an explicitly requested council of named models or agents with a judge. Ordinary review and brainstorming do not need a council. --- # Model Council Run the same task through user-selected independent members, then a selected judge. Preserve every member's full report and the judge's consolidation. Resolve members and judge from this request or its existing explicit selections; ask only for missing choices or unreachable members. Inspect current runtime configuration and the available permitted dispatch interfaces. Do not infer cross-provider access from a model name, put effort only in prose when a runtime control exists, or silently substitute a member. Follow shared delegation policy. For Codex members from Claude Code, call the Codex CLI directly (`codex exec`), never the retired codex-adapter plugin; see `AGENTS.md` §Running Codex. 1. Write one common brief: question, relevant source pointers, constraints, acceptance and report format. Use a neutral brief without preferred conclusions. 2. Give each member that brief independently. Parallelism is useful but optional; isolation means members do not see one another's answers. Save full attributed reports under the supplied run directory or `scratchpad/council/<topic>/`. 3. Give a fresh judge the brief and all attributed reports. Request a synthesis of agreement, disputes, evidence, best-supported conclusion and uncovered gaps. Majority agreement is not proof; the judge may favor a supported minority view. 4. Deliver a concise consolidation, member-position summaries and paths to every full report. Include full text inline when requested, rather than discarding it. Record actual model identities where the runtime provides evidence. A failed member gets bounded safe recovery; report any missing member and whether the judge saw an incomplete council. Never backfill a member with your own answer. A council-only request does not authorize implementing its recommendations.