Communitygithub.com

amplitude/builder-skills

设计带有正确指标和变体的A/B测试,分析进行中或已完成的实验,并以统计严谨性解释结果。在设置实验、检查实验状态、分析结果或做出发布决策时使用。

builder-skills 是什么?

该技能是一个全面的实验分析工具,专门用于Amplitude平台,帮助数据科学家、产品经理和工程师基于统计严谨性做出数据驱动的发布或不发布决策。它支持进行深入的实验分析,包括识别实验、检索和验证设置、检查数据质量(如样本量、统计功效、置信区间宽度以及七个有效性标志)、分析主要指标、次要指标和防护栏指标,并执行细分分析以揭示细分市场之间的异质效应。该技能还评估实验持续时间是否足够,将定量结果与定性反馈关联起来,并综合所有发现以提供明确的建议。它通过严格的统计框架确保分析结果具有可操作性,避免匆忙或肤浅的总结,使用具体的数字和业务影响解释来支持决策。

兼容平台~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/amplitude/builder-skills/tree/HEAD/analytics-skills/skills/analyze-experiment

在你喜欢的 AI 中提问

打开一个已预加载此 Agent Skill 的新对话。

文档

实验分析师

对实验进行全面、深入的详细分析,以做出数据驱动的发布/不发布决策。这不是快速总结——提供包含具体数字和业务影响的深入见解。

使用时机

  • 分析已完成的实验结果以做出发布决策
  • 检查进行中实验的进展和早期信号
  • 了解实验为何成功或失败
  • 调查意外结果或细分层面的效应

分析理念

全面而非简短:

  • 包含具体数字、百分比和数据点
  • 用通俗语言解释统计意义和业务影响
  • 覆盖所有指标(主要、次要、防护栏),提供实际值
  • 这是一次全面的分析——不要仓促行事或提供肤浅的总结

说明

步骤 0:识别实验

如果用户提供了特定实验:

  • 接受实验 URL 或实验 ID
  • 如果是 URL:使用 Amplitude:get_from_url 提取详细信息
  • 如果是 ID:进入步骤 1

如果用户一般性地询问实验:

  • 使用 Amplitude:search,参数为 entityTypes: ["EXPERIMENT"] 和相关查询词
  • 显示前 3-5 个匹配项,包括名称、ID 和状态
  • 询问用户要分析哪个实验

如果未指定实验:

  • 明确询问实验 URL、ID 或搜索词,然后停止

步骤 1:检索并验证设置

使用 Amplitude:get_experiments(传入实验 ID)获取:

  • 实验名称、键、描述和状态
  • 开始/结束日期和持续时间
  • 变体:名称、流量分配
  • 附加指标:主要(recommendation=true)、次要、防护栏(存储为 ID)
  • 分桶策略

获取指标名称:

  • 从实验响应中提取指标 ID(例如 "c4pn8fkv")
  • 关键:Amplitude MCP 无法通过 ID 直接检索指标名称
  • 替代方案:
    1. 使用 Amplitude:search,参数为 entityTypes: ["CHART"] 和实验名称,搜索与实验相关的图表
    2. 使用 Amplitude:get_charts 检查相关图表的定义以查找指标引用
    3. 检查实验描述是否包含指向指标文档的链接
  • 如果无法找到指标名称,则报告为描述性占位符:
    • 主要指标:“主要目标指标 (ID: {id})”
    • 次要指标:“次要指标 {index} (ID: {id})”
    • 包含指标 ID,以便用户可以在 Amplitude UI 中查找

验证:

  • 实验是否正在运行或已完成?(而非草稿)
  • 是否已运行 1 周以上?
  • 是否明确定义了变体和指标?

如果不完整,解释缺失了什么并停止。


步骤 2:检查数据质量(使用明确的阈值)

使用 Amplitude:query_experiment(仅主要指标)评估:

流量平衡(SRM 检查):

  • 报告每个变体的实际流量分配(例如,48.2% 对照组,51.8% 处理组)
  • 使用 API 中的 srmDetected 字段: 如果 srmDetected: true 则标记
  • SRM(样本比例不匹配)表示观察到的流量分配与预期分配存在显著偏差
  • 如果检测到 SRM,报告预期与实际分配的具体百分比
  • 严重的 SRM 可能指示检测工具问题或分桶问题,可能使结果无效

样本量分析:

A. 当前样本评估:

  • 报告每个变体的总用户数,包含具体数字
  • 如果每个变体 <100 个用户则标记(不足以得出任何结论)
  • 如果用户数为 100-1000 则标记(仅方向性信号,不能做出自信决策)
  • 每个变体需要 1000 以上才能做出自信决策

B. 统计功效分析:

  • 目标效应量: 对业务而言有意义的最小提升是多少?(通常为转化指标的 2-5%)
  • 实现功效: 给定当前样本量和观察到的方差,检测到目标效应(如果存在)的概率是多少?
  • 功效解释:
    • <50%:严重功效不足——可能错过真实效应
    • 50-70%:功效不足——假阴性风险高
    • 70-80%:勉强足够——如果 p 值处于临界值,考虑延长
    • 80%+:功效良好——足以检测到目标效应量
  • 如果功效不足: 计算达到 80% 功效所需的额外样本量
  • 建议: 如果功效 <70% 且结果不确定,延长持续时间而非做出仓促决策

C. 精度分析(置信区间宽度):

  • 主要指标的 CI 宽度: 将 95% 置信区间宽度报告为基线的百分比
  • 精度评估:
    • CI 宽度 > 基线的 10%:低精度——效应量不确定性太高,无法做出自信决策
    • CI 宽度为基线的 5-10%:中等精度——可接受用于方向性决策
    • CI 宽度 < 基线的 5%:高精度——足够窄,可以做出自信决策
  • 可操作性阈值: CI 是否能窄到区分实际显著效应和可忽略效应?
    • 如果 CI 下限提示有意义的提升但上限边缘,则精度可能不足
    • 示例:如果目标是 +5% 提升,CI 为 [-2%, +12%],则太宽,无法自信地断定效应超过目标
  • 建议: 如果 CI 太宽,延长持续时间或增加流量分配以提高精度

综合数据质量标志:

Amplitude:query_experiment API 返回多个布尔标志,用于评估统计有效性。检查并记录每个:

  1. statsAssumptionsMetForWholeExperiment:

    • 表示是否满足核心统计假设(正态性、独立性)
    • 如果为 false: 结果可能不可靠;考虑非参数方法或更长的运行时间
    • 影响:高——影响所有统计结论
  2. hasSuspiciousUplift:

    • 标记意外大的效应量,可能指示数据质量问题
    • 如果为 true: 验证检测工具,检查机器人流量或细分异常
    • 影响:高——可能表示测量错误而非真实效应
  3. isVariancePositive:

    • 确认指标方差为正(统计检验在数学上所需)
    • 如果为 false: 关键数据质量问题——指标可能是常数或计算错误
    • 影响:关键——如果为 false,统计检验无效
  4. isConfidenceIntervalNotFlipped:

    • 确保 CI 下限 < CI 上限(数学一致性检查)
    • 如果为 false: 表示计算错误或数据损坏
    • 影响:关键——结果不可信
  5. isStandardErrorLargeEnough:

    • 检查标准误差是否足够大以进行可靠推断
    • 如果为 false: 高方差或非常小的样本可能产生不可靠的置信区间
    • 影响:中等——影响估计精度
  6. isPointEstimateInsideConfidenceInterval:

    • 验证点估计是否落在其置信区间内(一致性检查)
    • 如果为 false: 计算错误或数值不稳定
    • 影响:高——表示统计计算问题
  7. isMeanValid:

    • 确认均值是有效数字(不是 NaN,不是无穷大)
    • 如果为 false: 数据质量问题——检查空值或计算错误
    • 影响:关键——如果均值无效则无法分析

对于每个失败的标志(返回 false 或对于 suspicious uplift 返回 true),记录:

  • 哪个标志失败
  • 用通俗语言解释其含义
  • 对结果可靠性的具体影响
  • 建议的操作(延长持续时间、调查检测工具等)

如果所有标志都通过: 明确记录为强数据质量信号

时间稳定性:

  • 检查主要指标是否日复一日地稳定
  • 注意斜坡期(前 24-48 小时)或一周中的某天效应

记录发现的所有数据质量问题——这些会影响结果的可靠性。


步骤 3:分析主要指标

使用 Amplitude:query_experiment 不带 metricIds 仅获取主要指标。

使用步骤 1 中的指标名称——报告时使用人类可读的指标名称,而非指标 ID。

提取并报告:

  • 对照组基线: 指标值和样本量
  • 处理组表现: 指标值和样本量
  • 绝对提升: 处理组 - 对照组
  • 相对提升: (处理组 - 对照组) / 对照组 × 100%
  • P 值: 附解释
  • 置信区间: 报告 95% CI 边界

解释:

  • ✅ 统计显著: p < 0.05 且 CI 不包含 0
  • ⚠️ 趋势性: 0.05 < p < 0.15(有提示但不确定)
  • ❌ 无效应: p ≥ 0.15 或 CI 包含 0

实际显著性:

  • 提升幅度对业务是否有意义?
  • 即使统计显著,小的提升(<2-3%)可能不值得增加复杂性
  • 考虑指标的业务影响(收入 vs. 低价值参与)

步骤 4:分析次要指标和防护栏

使用 Amplitude:query_experiment 带 metricIds 用于所有指标。

使用步骤 1 中的指标名称——报告时使用人类可读的指标名称,而非指标 ID。

对于每个次要指标:

  • 报告指标名称(来自步骤 1 的映射)、变体表现和统计显著性
  • 注意哪些移动了,哪些没有(包含具体数字)
  • 识别意外后果: 标记任何负面影响的特定值

对于每个防护栏:

  • ✅ 无回归:中性或正面(p > 0.05)
  • ⚠️ 边缘问题:小的负面提升(1-5%)且 p < 0.10
  • 🚩 显著回归: 负面提升且 p < 0.05——报告实际数字

关键问题: 是否有任何指标显示退化(收入、留存、参与、错误率)?

多重检验: 如果分析 5 个以上指标,考虑 Bonferroni 校正(alpha = 0.05 / 指标数量)


步骤 5:全面细分分析

使用 Amplitude:query_experiment 并带 groupBy 参数(一次一个)。

测试 3-4 个高信号细分:

  1. 平台(iOS、Android、Web)
  2. 用户使用时长(新用户 vs. 老用户)
  3. 计划类型(免费 vs. 付费)
  4. 地理位置(国家、地区)

强制要求:将结果格式化为 Markdown 分解表格

对于每个细分分析,按以下精确格式呈现结果:

细分对照组率对照组曝光数对照组占比处理组率处理组曝光数处理组占比相对提升显著?
iOS48.7%1,23445.2%55.4%1,45654.8%+13.6%是 (p=0.02)
Android63.9%56720.8%65.1%58922.2%+1.9%否 (p=0.45)
Web51.2%92834.0%50.8%61123.0%-0.8%否 (p=0.89)

计算占比:

  • 将所有细分中的曝光数相加得到总数
  • 显示每个细分的份额:(细分曝光数 / 总曝光数) × 100%
  • 这揭示哪些细分推动了整体结果

关键见解:

  • 识别处理组表现 最佳 的细分(定向发布机会)
  • 识别处理组 有害 的细分(考虑排除)
  • 解释为什么不同细分表现出不同性能
  • 注意辛普森悖论: 整体结果可能与所有细分结果不同

使用 groupByLimit: 10 以避免输出过多。


步骤 6:评估持续时间和运行充分性

持续时间评估: 基于步骤 2 中的功效和精度分析,评估实验是否运行了足够长的时间:

运行时因素:

  • 最小时长: 实验是否至少运行了 1-2 周以捕获完整的用户生命周期?
  • 学习效应: 对于功能变更,用户是否有时间适应?(通常 3-7 天)
  • 每周季节性: 实验是否捕获了至少一个完整周以考虑一周中的某天模式?
  • 业务周期: 对于 B2B 产品,是否已运行过完整的业务周模式?

与步骤 2 的功效分析集成:

  • 如果步骤 2 显示足够功效(>80%)且 p < 0.05: 实验有足够数据,持续时间充足
  • 如果步骤 2 显示低功效(<70%)且 p > 0.05: 数据不足导致不确定,延长持续时间
  • 如果步骤 2 显示足够功效(>80%)且 p > 0.15: 数据足以接受零结果(无效应)
  • 如果步骤 2 显示足够功效但 CI 宽度太宽: 需要更多数据以提高精度,延长持续时间

速度预测(仅在建议延长时):

  • 当前每日注册: 计算每个变体每天的用户数
  • 达到目标样本所需天数: 基于步骤 2 的功效计算,还需要多少天?
  • 达到目标精度所需天数: 基于步骤 2 的 CI 宽度计算,还需要多少天才能达到所需精度?
  • 建议: 提供实验应达到足够功效/精度的具体日期

不要重复步骤 2 中的功效计算——引用那些发现,重点放在持续时间和时间表建议上。


步骤 7:理解原因(定性背景)

对于显著结果(正面或负面):

使用 Amplitude:get_feedback_insights:

  • 按实验日期范围过滤
  • 对于成功:查找 ["lovedFeature", "mentionedFeature"]
  • 对于失败:查找 ["bug", "complaint", "painPoint"]
  • 检查主题是否与实验假设一致

将定性与定量联系起来:

  • 用用户引述或反馈主题解释提升
  • 呈现 2-3 个代表性示例,包含具体细节

步骤 8:综合发现并提出建议

在最终确定之前,请验证您是否已包含:

  • ✓ 所有主要指标数据(提升、CI、p 值、解释)
  • ✓ 所有数据质量发现(SRM、样本量、功效、精度、所有 7 个有效性标志及其实际值)
  • ✓ 所有次要指标和防护栏(包含实际值和显著性)
  • ✓ 所有细分分析表格(格式化为含曝光占比)
  • ✓ 统计功效评估(当前功效、所需样本、持续时间)
  • ✓ 定性背景(如适用)

明确建议:

发布 🚀

  • 条件:主要指标显著改善,防护栏无退化,数据质量强,细分分析正面
  • 附带:发布策略建议(全面发布、阶梯式推出、定向发布)

不发布 ❌

  • 条件:主要指标负面或零结果且足够功效,防护栏退化,数据质量问题
  • 附带:下一步建议(迭代想法、不同假设、调查数据问题)

不确定 / 需要更多数据 ⏳

  • 条件:趋势性结果,功效不足,CI 太宽,细分分析混合
  • 附带:建议(延长持续时间,增加流量分配,调整指标)

步骤 9:提供后续步骤

  • 对于发布:建议监控期和回滚标准
  • 对于不发布:建议根本原因调查或替代假设
  • 对于更多数据:建议具体行动(例如,“在 40% 流量下再运行 14 天以达到 80% 功效”)
  • 提供调试帮助:SQL 查询、图表链接或 Amplitude MCP 操作以进一步调查

始终包含具体数字——从不使用“大部分”、“少数”或“许多”。如果进行多个查询,使用单独的 query_experiment 调用。

Individual skills in this repo

This repo contains 20 individual skills — each has its own dedicated page.

amplitude/builder-skills

Performs deep analysis of a specific Amplitude chart to explain trends, anomalies, and likely drivers. Use when a metric looks unusual, investigating a spike or drop, or understanding the "why" behind numbers.

amplitude/builder-skills

Deeply analyze Amplitude dashboards by analyzing key charts, surfacing top areas for concern and takeaways, identify anomalies, then explain changes using customer feedback trends.

amplitude/builder-skills

Synthesizes customer feedback into actionable themes including feature requests, bugs, pain points, and praise. Use when planning product roadmap, understanding user sentiment, investigating specific issues, or preparing voice-of-customer reports.

amplitude/builder-skills

Analyze MCP server usage instrumented with Amplitude's MCP Analytics SDK: break usage and errors down by tool, read the rationales within each tool to see what callers are trying to do, and produce a prioritized write-up of actionable fixes. Use this skill whenever the user asks to understand how their MCP server is being used, what agents/users are trying to do with it, why tool calls are failing, what to fix or improve in their MCP server, or asks for an "MCP usage report", "tool error analysis", "intent analysis", "rationale clustering", or "MCP insights". Also trigger when the user mentions [MCP]-prefixed events, tool rationale, tool call errors, or just finished instrumenting their MCP server and wants to see what the data says. Requires the Amplitude MCP connector.

amplitude/builder-skills

Read lost deals and churned accounts from your CRM, extract reasons clustered by theme (missing features, pricing, competitors, UX), and write a prioritized weekly analysis with product improvement recommendations. Use before roadmap planning or to build the case for prioritizing retention work.

amplitude/builder-skills

Creates Amplitude charts from natural language descriptions, handling event selection, filters, groupings, and visualization choices. Use when you know what you want to measure but prefer not to build the chart manually.

amplitude/builder-skills

Guide an Amplitude user through building a custom agent by suggesting use cases grounded in their role and data, shaping the idea into a well-formed spec, and generating a ready-to-run Global Agent deeplink that creates it. Use to create, build, or set up a custom agent, automate a recurring analysis, or put a repeated report on a schedule.

amplitude/builder-skills

Builds comprehensive Amplitude dashboards from requirements or goals, organizing charts into logical sections with appropriate layouts. Use when creating a complete dashboard from scratch or assembling existing charts into a cohesive view.

amplitude/builder-skills

Monitors all active and recently completed experiments across Amplitude projects, triages them by importance, then runs deep analysis and reporting on the most impactful ones. Use when the user asks to "check on experiments", "experiment status", "experiment review", "what experiments are running", or wants a periodic experiment health report.

amplitude/builder-skills

Pull Intercom tickets and Slack support messages from the past 7 days, classify each signal, enrich with CRM data (ARR, plan, renewal), score by customer value and churn risk, and output a tiered priority report saved to Drive. Use when you need a fast, data-driven view of what support signals matter most.

amplitude/builder-skills

Use this skill whenever a user wants to improve existing pages on their website to get cited more by AI models — whether they say "our pages aren't getting cited", "improve this page for AI visibility", "which of our pages should we update", "make this article more cite-worthy", "our competitors are getting cited instead of us", "update our content for AI search", or any variation where the goal is improving an existing asset rather than creating something new. This skill pulls owned pages from AI Visibility, identifies which ones have citation potential but are underperforming, compares them against the external pages that are winning citations on the same topics, and produces section-level rewrites or a full-page update — then pushes the revision to the CMS as a draft. Trigger even if the user just says "help me get cited more" or "why is [competitor] getting cited instead of us".

amplitude/builder-skills

Use this skill whenever a user wants to win AI citations on prompts that competitors currently dominate — whether they say "competitors are getting cited instead of us", "we're losing on these prompts", "how do I outrank [competitor] in AI answers", "find prompts where we should be winning", "create content to beat [competitor]", or any variation where the goal is capturing AI share on prompts a competitor currently owns. This skill pulls competitor visibility data from AI Visibility, identifies the specific prompts where competitors win and Amplitude is absent, clusters them by intent, and produces targeted comparison pages, alternatives content, or rebuttal assets — then pushes drafts to CMS. Trigger on any mention of competitor, prompt hijack, outrank, or "why is [competitor] getting cited instead of us".

amplitude/builder-skills

Use this skill whenever a user wants to turn AI Visibility data into published content — whether they say "find content gaps", "what should we write about", "which topics have low visibility", "help me get cited by AI models", "create a blog post from our AI Visibility gaps", "we're losing to competitors on these prompts", or any variation where they want to go from AI visibility weakness to a draft article, landing page, or FAQ. This skill connects directly to Amplitude AI Visibility data (topics, prompts, visibility scores, citations, competitor data, full LLM responses and sources) and produces a publish-ready content brief plus full article draft. If the user mentions CMS (WordPress, Webflow, Contentful, Sanity, HubSpot, Ghost, Shopify), also trigger this skill to push the draft directly. Trigger even if they just say something vague like "what content should we create?" in an AI Visibility context.

amplitude/builder-skills

Use this skill whenever a user wants to test content variants before publishing to find which one will get cited most by AI models — whether they say "which version of this content will perform better", "test this article before we publish", "simulate how AI will respond to this content", "which angle should we use", "generate content variants and pick the winner", "run a simulation before publishing", or any variation where the goal is data-driven content selection rather than gut-feel publishing. This skill takes an identified content opportunity, generates 2–3 distinct variants with different angles or structures, scores them against actual AI model responses from AI Visibility, references the Simulate Changes feature for pre-publish validation, and produces a clear recommendation on which variant to publish — then pushes the winner to CMS. Trigger on any mention of "simulate", "test variants", "which performs better", "A/B content", or "before we publish".

amplitude/builder-skills

Use this skill whenever a user wants to understand which external sources are being cited by AI models on topics relevant to their brand, and wants to create content that will outrank those sources — whether they say "what sources are AI models citing", "why is [third-party site] being cited instead of us", "we want to be the definitive source on X", "build something that gets cited more than G2 or TechRadar", "create an authoritative asset", or any variation where the goal is producing a new reference asset (definition page, benchmark, methodology, glossary, comparison hub) designed to beat existing top-cited sources. This skill analyzes AI Visibility source data, reverse-engineers what makes top-cited pages authoritative, and produces a superior source asset — then pushes it to CMS as a draft. Trigger on any mention of "sources", "third-party citations", "authoritative content", "definitional pages", or "outrank".

amplitude/builder-skills

Instrument a Node/TypeScript MCP server with Amplitude's @amplitude/mcp-analytics SDK so tool calls, sessions, and rationale are tracked as Amplitude events. Use this skill whenever the user wants to add Amplitude analytics to their MCP server, mentions "MCP Analytics", "@amplitude/mcp-analytics", "instrument my MCP server", "track MCP tool calls", "add rationale to my MCP tools", or wants agent traffic (Claude, Cursor, ChatGPT) attributed back to Amplitude. Also use for adding UTM tagging to MCP-returned links, or for troubleshooting identity/user_id mismatches between MCP events and web/mobile Amplitude data.

amplitude/builder-skills

Instruments a pull request with Amplitude analytics that conform to the project's existing taxonomy. Reads the tracking plan via the Amplitude MCP server (events, properties, naming conventions), analyzes the PR diff to find the few user actions genuinely worth tracking, detects the codebase's SDK and tracking patterns, and adds instrumentation that matches both. Optionally (opt-in) stages new events and properties on an Amplitude tracking-plan branch for data-governance review. Use when asked to "instrument this PR", "add analytics to this change", "add tracking", "add Amplitude events", "instrument this feature", or "what should I track here".

amplitude/builder-skills

Diagnoses product health by cross-referencing Amplitude analytics (dashboards, charts, funnels, feedback, AI agent analytics), optionally Datadog (errors, latency, stack traces), and optionally Slack (qualitative feedback, bug reports, feature requests). Identifies what's broken, what's working, and what to do about it — with root causes, not just symptoms. Use when asked to "diagnose my product", "what's going on", "product health check", "what's broken", "where are users struggling", "give me a product diagnosis", or "what should I focus on".

amplitude/builder-skills

Turn one or more meeting transcripts, notes, or Slack threads into concise takeaways and clear action items with DRIs. Works with a single meeting or a batch from the whole week.

amplitude/builder-skills

Summarizes B2B account health by analyzing usage patterns, engagement trends, risk signals, and expansion opportunities. Use for customer success reviews, renewal preparation, QBRs, or account prioritization.

相关技能