More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    终极圆桌:六专家对抗式决策引擎

    1

    让六个 AI 专家互相拷问你的决策,最后逼出一个带负责人、KPI 和止损线的决定。

    Free

    Security scanned
    终极圆桌:六专家对抗式决策引擎

    终极圆桌:六专家对抗式决策引擎

    Example session with this skill installed

    开启终极圆桌. Topic: Whether to switch from annual to quarterly performance reviews. North Star: Monthly Active Users (MAU). memory_store: .ultimate/memories.jsonl.

    • Read your context and instructions
    • Compiled the output

    The roundtable has concluded following a semantic deadlock trigger in Stage 2. The participants have voted on the final strategy to prioritize operational efficiency for MAU growth.

    ⚔️ 终极圆桌决策报告 V4.1

    📌 1. 北极星指标与底线假设
    • 本次推演的核心指标:MAU (Monthly Active Users)
    • 我们击穿的隐性假设:年度评审导致的周期性士气低落对 MAU 的负面影响可以通过管理手段对消。
    🔗 2. 证据链与外部验证
    • 工具验证结果:Search 工具确认 SaaS 行业采用触发式评审的团队,其研发效能波动率降低 12%。
    🚦 3. 铁血裁决(三选一)
    • 激进方案 A:全量改为季度评审,三个月内切换完毕(投票支持:musk)
    • 保守方案 B:维持年度评审,仅精简一半材料流程(投票支持:munger, pg)
    • 降维方案 C:取消固定评审,改为触发式评审:指标跌破阈值才开(投票支持:feynman, jobs, karpathy)
    • 🏆 最终拍板结论:采用方案 C
    🎯 4. 战术级行动清单 (Action Items)
    • 负责人:karpathy
    • 第一步(Start):本周在 2 个核心增长团队上线触发式评审系统,预警阈值设为 MAU 跌破 8%。
    • 关键指标(KPI):30 天内触发次数控制在 5 次以内,且期间 MAU 环比增长不低于 9%。
    • 止损红线(Kill):如果 30 天内触发超过 12 次,立刻停止该机制并回滚到年度评审模式。
    • 补充调研(Need Info):过去两年每次评审的平均人力耗时。
    🧠 5. 长期记忆写入
    • [MEMORY_UPDATE]:确认组织架构调整应优先于流程调整,触发式评审已作为标准化风险对冲方案入库。

    Next steps

    • Deploy the deadlock_detector.py to archive the Stage 2 debate transcripts.
    • karpathy to execute the monitoring script for the 2 pilot teams.
    • Review the Need Info data in the next 72 hours.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    大多数多智能体讨论技能,产出的是六份互相恭维的备忘录——没人被逼着证明什么,最后什么也没决定。

    这个技能让六个人互相拷问,并且只允许一种结尾:一个决定。

    【四个真正起作用的东西】

    1. 拒绝角色扮演 每个角色跑的是强制思维协议,不是人设服装。马斯克必须拆到物理约束并给出量级估算;费曼必须用外行话讲清否则标记"未理解";芒格必须先反向问"怎么才能保证它失败";乔布斯必须回答"这更简单还是更复杂";Karpathy 必须让每条论断可复现;Paul Graham 必须从真实案例起手。模糊表述在校验阶段就被拒绝,不会留到最后。

    2. 死锁熔断 内置语义级死锁检测。当双方只是在换说法复读(同样的论点、不同的词、不同的例子),系统切断辩论、强制进入裁决。逐字重复容易抓,换说法才是难点,这里抓的就是后者。

    3. 可执行的产出,不是摘要 收尾只有三样东西:三个互斥方案(激进 / 保守 / 降维)、六人强制投票、一个决定。然后是一张命令单:负责人、第一步、带数字的 KPI、带数字的止损线。写"如果效果不好就停止"会被校验器拒绝——模糊是一种构建错误。

    4. 零摩擦 附带一键跑通与交互式向导。python tools/quickstart.py:体检 → 跑完一场完整讨论 → 自检 → 下一步,约 10 秒,无需配置、无需 API key、无依赖。python tools/wizard.py:交互式生成你自己的五个阶段配置,边填边校验。

    【还包含】

    · 诚实标注:每个数据点都标为真实工具或模拟数据,报告头部直接印出比例。缺标注会被拒绝——模拟数据不能冒充真实。 · 注意力分舱:每个角色只看到与自己视角相关的历史,避免上下文过载变成一锅粥。 · 少数派保护:未被有效反驳的少数派观点完整保留,不会被多数票淹没。 · 平民视角审计:六人全票通过时提醒你——盲区通常属于那个真正要用它的人。 · 跨会话记忆:经验在多次运行之间沉淀(纯 JSONL 文件)。

    【安装】

    把文件夹放进你的 agent skills 目录: Claude Code ~/.claude/skills/ultimate-roundtable/ Cursor / Cline <项目>/skills/ultimate-roundtable/ 马维斯 技能管理 → 导入技能 → 选 SKILL.md 其他(读 AGENTS.md 的) <项目>/skills/ultimate-roundtable/

    然后对它说:「开启终极圆桌:要不要把年度评审改成季度评审」

    【自定义宿主环境】

    如果你的宿主不支持 ZIP 导入,克隆 GitHub 仓库后运行 python tools/quickstart.py,30 秒即可。

    【要求】

    Python 3.9+,仅用标准库,无需安装任何依赖。 支持 Claude Code、Codex、Cursor、Cline、Windsurf,以及任何读取 SKILL.md 的 agent。

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Agents & Orchestration.

    Requires a North Star metric. Semantic detection accuracy depends on the underlying LLM capability. Fails if action items lack numeric KPIs or owners.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Free to use

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 16 days ago

    • Free with an Agensi account

    Software
    Python 3.9+

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Environment Variables

    This skill is fully local and offline by default. • No network requests — the scripts use Python standard library only (no urllib/requests/socket anywhere). • No browser automation — no selenium/playwright/webbrowser. • No third-party accounts, API keys, or OAuth. Nothing to connect. Why each permission is needed: • Terminal/Shell — the orchestrator is a Python script; every stage is a CLI call. • Read/Write Files — it maintains a discussion state file (state.json) and writes the final report. • Environment Variables — one optional override, ROUNDTABLE_MEMORY, lets you point the cross-session memory store at a custom path. If unset, it defaults to ./.roundtable-memory/ inside your project. File scopes are limited to the project workspace: .roundtable/ (state), .quickstart/ (demo output), and optionally ~/.marvis-roundtable/ only when the project is not a git repo. Python 3.9+ required. Nothing to pip install.

    Listed16 days ago

    What's inside

    Frequently Asked Questions

    This skill is designed to end 'analysis paralysis' by using a rigid state machine that detects repetitive arguments and forces experts to choose between three concrete, mutually exclusive strategic paths.

    You will receive the 'ultimate_orchestrator.py' file, the local JSON-based state machine framework, and the specialized scripts required for semantic deadlock detection and attention compartmentalization.

    The skill requires a Python environment to run the orchestrator and works best with agents that support structured data output; it also requires you to define a single 'North Star' metric before initialization.

    The skill includes a 'degraded mode' to ensure functionality if specific role-logic is missing, but for optimal results, we recommend using high-context LLM models that can handle the 150-character limit constraints.

    Unlike standard prompts that often produce vague 'consultant speak,' this skill uses a hard-coded execution list that refuses to finalize a report unless it includes numeric KPIs and a project 'kill switch.'

    New in Agents & Orchestration

    Free