agentsclimarketplace

Experiment orchestrator

Skill LuckyOneTwoThree/pm-skill/pm-05-metrics-ops/orchestrators/experiment-orchestrator

当需要设计或执行A/B测试实验时使用。实验验证指挥官,调度experiment-design/execution。关键词:A/B测试、实验设计、统计显著性、实验执行、效果验证、AB测试、对照实验。From its SKILL.md

Install
npx -y skills add LuckyOneTwoThree/pm-skill --skill experiment-orchestrator

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
  • 6 stars6 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

7.2 KB, ~2.3k tokens by cl100k_base, as published. Nobody here has run it

实验设计指挥官

核心原则

实验是学习的最快方式

每一个实验都是一次有控制的探索,目标不是证明假设正确,而是以最快速度获得可靠的学习。实验的价值在于学习速度,而非实验数量。

编排理念

  1. 设计→执行两阶段缺一不可:没有设计的执行是盲目的,执行阶段包含结果分析和报告生成
  2. 人类审核是实验的必要卡口:实验方案和实验报告都必须经人类审核,执行过程可自动化
  3. 护栏指标一票否决:无论主指标多正向,护栏指标突破即暂停

编排协议

编排协议遵循 orchestrator-protocol.md 统一标准。

Pipeline

pipeline: experiment-orchestrator
version: 8.0

post_pipeline:
  - action: stage-summary
    output: output/phase-reports/pm-metrics-ops/experiment-orchestrator.md

stages:
  - id: phase-1
    name: "实验设计"
    depends_on: []
    skills: [experiment-design]
    gate:
      condition: "实验设计经人类审核确认"
      fail_action: "阻止实验上线,修改后重新审核"

  - id: phase-2
    name: "实验执行"
    depends_on: [phase-1]
    skills: [experiment-execution]
    gate:
      condition: "样本量充足且统计检验完成、实验报告经人类审核确认"
      fail_action: "延长实验周期或扩大流量"

阶段执行计划

调用 experiment-design

Skill: experiment-design
输入:
  hypothesis: 用户提供(假设陈述)
  available_traffic: 用户提供(可用流量)
  metrics_system: metrics-system → metrics.json(可选)
  historical_data: analysis-funnel/analysis-retention(可选)
输出: output/pm-metrics-ops/experiment-design/
验证: 假设已结构化(If-Then-Because-For);主指标与假设直接对应;护栏指标覆盖留存、收入、技术三个维度;样本量计算参数有据可依
模式: 🤖→👤

调用 experiment-execution

Skill: experiment-execution
输入:
  experiment_design: output/pm-metrics-ops/experiment-design/experiment_design.json
  experiment_data: 用户提供
  termination_conditions: output/pm-metrics-ops/experiment-design/experiment_design.json
  product_background: 用户提供(可选)
输出: output/pm-metrics-ops/experiment-execution/
验证: 实验分组流量分配正确;护栏指标未触发告警;实验数据采集完整;统计显著性计算正确;统计结论与数据一致;行动建议与结论一致;护栏指标全覆盖;异质性效应已分析(至少3个分群维度)
模式: 🤖→👤

阶段总结(post_pipeline)

所有业务阶段执行完成后,必须立即生成阶段总结文档:

动作: 生成阶段总结
输入:
  所有子Skill输出: output/pm-metrics-ops/
  人类决策记录: 本轮执行中的人类决策点及结果
输出: output/phase-reports/pm-metrics-ops/experiment-orchestrator.md
验证: 阶段总结文档已生成,6项结构(执行概览/关键发现/决策记录/产出清单/风险与待办/下游衔接)均非空
下游衔接:
  primary:
    target: decision-orchestrator
    reason: 实验完成,将实验结论转化为决策行动
    input_mapping:
      experiment_output: "output/pm-metrics-ops/experiment-execution/ → decision-dace输入"
  alternatives:
    - target: release-orchestrator
      reason: 实验结果显著,建议全量发布
      condition: 实验结果统计显著(p<0.05)且业务意义达标时
    - target: analysis-orchestrator
      reason: 实验结果需更深入的数据分析
      condition: 实验结果存在异常或需多维下钻时
  special_cases: []
模式: 🤖

⏸ 阶段卡口:阶段总结文档已生成且6项结构均非空 → 未通过:补充缺失结构项后重新生成

阶段卡口

卡口条件未通过处理
实验方案人类已审核实验设计经人类审核确认阻止实验上线,修改后重新审核
统计显著性已判断experiment-result输出文件已生成且非空延长实验周期或扩大流量
实验报告已审核实验报告经人类审核确认补充分析或修改结论
阶段总结已生成output/phase-reports/pm-metrics-ops/experiment-orchestrator.md 已生成且6项结构均非空补充缺失结构项后重新生成

人类决策点

决策点触发条件决策内容
实验方案审核实验设计完成审核假设合理性、指标选择、分流方案
全量/终止决策实验结果分析完成决定全量发布、终止实验或延长周期
实验报告确认实验报告生成完成确认报告结论和行动建议

决策规则

条件Action
样本量达到100%立即触发结果分析
统计显著(p < 0.05)且稳定考虑提前终止
护栏指标显著下降触发告警,考虑终止
新奇效应显著延长实验周期
实验组持续负向考虑提前终止

异常处理

异常类型处理策略
实验设计人类审核未通过阻止实验上线,返回设计阶段修改,不进入执行阶段
护栏指标突破阈值立即暂停实验执行,触发告警,提交人类决策是否终止实验
实验数据采集异常标记数据异常,暂停统计检验,提示人类检查数据管道
实验报告人类审核未通过返回执行阶段补充分析,不传递到下游
多实验流量冲突按优先级排队,低优先级实验暂停,标注"流量冲突"
阶段总结生成失败基于已完成的子Skill输出生成部分总结,缺失项标注"数据缺失",不阻塞编排完成

变更记录

  • v1.0: 初始版本
  • v2.0: 结构优化
  • v3.0: 新增 experiment-report(A/B测试报告)
  • v4.0: 编排器优化——任务调度改为阶段执行计划,新增子Skill执行协议,调度规则改为执行模式,阶段卡口和人类决策点改为表格
  • v5.0: 执行步骤原则替换为编排理念,新增异常处理表
  • v6.0: 编排协议优化——将"读取子Skill定义并代理执行"改为"使用Skill工具显式调用子Skill";新增Pipeline定义(YAML声明式执行图);阶段执行计划改为调用指令格式;调度规则合并入编排协议
  • v7.0: 阶段总结强化——Pipeline新增post_pipeline定义;调用规则第6条改为强制执行;阶段执行计划新增阶段总结执行指令;阶段卡口新增阶段总结校验;异常处理新增阶段总结生成失败策略
  • v8.0: 合并 experiment-report 到 experiment-execution,移除 experiment-report 阶段,更新编排理念为设计→执行两阶段

What ships with it

Read from the repository

Just SKILL.md. No reference files, no scripts.

Keep looking

Skills are one crate of 625,857. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.