ab-test-analysis
Skill8 starsChanged 4 days agoMITView on GitHub
Use when designing, checking, or interpreting A/B tests and experiments. Computes significance, confidence intervals, sample ratio mismatch, power, CUPED, and sample size with a bundled zero-dependency script, then judges validity risks, guardrails, and ship/iterate/rollback decisions and renders an HTML experiment report. 中文触发词:A/B 实验、AB 测试、实验解读、显著性、样本量计算、能不能全量。
15 other projects use a version of this.
Adopted by
Most recently taken up by
- About-Intelligence/awesome-marketing-skillstheir own version, first seen Nov 8, 2026
- yuusakuri/agent-skillstheir own version, first seen Nov 1, 2026
- vijaysuperpositionist/skillstheir own version, first seen Nov 1, 2026
- ThinkingAIAgenticEngine/scenario-skillstheir own version, first seen Oct 17, 2026
- The-Utopia-Studio/skillstheir own version, first seen Oct 13, 2026
- Zafer-Liu/Data-Analysis-Agenttheir own version, first seen Oct 8, 2026
Used with
Projects with it also use brainstorm-experiments-new, brainstorm-experiments-existing and 10 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
skytiger6724Skillbrainstorm-experiments-new
Design lean startup experiments (pretotypes) for a new product. Creates XYZ hypotheses and suggests low-effort validation methods like landing pages, explainer videos, and pre-orders. Use when validating a new product idea, creating pretotypes, or testing market demand.
Also in 7 projects that use this
skytiger6724Skillbrainstorm-experiments-existing
Design experiments to test assumptions for an existing product — prototypes, A/B tests, spikes, and other low-effort validation methods. Use when validating assumptions, testing feature ideas cheaply, or planning product experiments.
Also in 5 projects that use this
45ckSkillpestle-analysis
Also in 5 projects that use this
launchapp-devSkillinterview-script
Create a structured user interview script with warm-up, core exploration, and wrap-up sections. Use when preparing for user research interviews to ensure consistent, insightful conversations.
Also in 7 projects that use this
AnthropicSkillsql-queries
Write correct, performant SQL across all major data warehouse dialects (Snowflake, BigQuery, Databricks, PostgreSQL, etc.). Use when writing queries, optimizing slow SQL, translating between dialects, or building complex analytical queries with CTEs, window functions, or aggregations.
Also in 7 projects that use this
nandanosqlSkillswot-analysis
Conduct a structured SWOT analysis — Strengths, Weaknesses, Opportunities, Threats — that connects internal capabilities to external forces for strategic decision-making.
Also in 8 projects that use this
- +6 more from skytiger6724
Other versions15 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 15
- Zafer-Liu/Data-Analysis-Agent★ 2.6ksame name
- jshsakura/awesome-opencode-skills★ 30same name
- henrique-simoes/Istara★ 15same name
- About-Intelligence/awesome-marketing-skills★ 14same name
- The-Utopia-Studio/skills★ 6same name
- ThinkingAIAgenticEngine/scenario-skills★ 6same name
- Adhamxon/claude-code-skills★ 2same name
- risadams/ink-and-agency★ 2same name
- skytiger6724/qwen-skills★ 2same name
- yuusakuri/agent-skills★ 2same name
5 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository23 other files it gives its agents.
- business-root-cause-analysisSkill
- daily-report-writerSkill
- dashboard-reviewerSkill
- data-analysis-report-writerSkill
- data-doc-writerSkill
- data-incident-postmortem-writerSkill
- data-lineage-impact-analyzerSkill
- data-presentation-architectSkill
- data-quality-rule-generatorSkill
- data-requirement-clarifierSkill
- data-tool-integration-plannerSkill
- data-viz-kitSkill
- exploratory-data-analysisSkill
- feature-engineering-advisorSkill
- funnel-analysisSkill
- market-research-analystSkill
- metric-definition-reviewerSkill
- model-evaluation-reviewerSkill
- retention-cohort-analysisSkill
- sql-reviewerSkill
- stream-pipeline-designerSkill
- table-design-advisorSkill
- weekly-monthly-report-writerSkill
The file
Files
A/B Test Analysis
目标
设计、审查或解读 A/B 实验,判断实验是否可信、效果是否真实、业务上是否值得推广。
核心原则:数字由脚本算,判断由人做。显著性、置信区间、SRM、功效和样本量全部用 scripts/ab_test.py 计算,不心算、不估算;模型负责检查实验设计、解释结果、补充业务判断。
使用场景
- 实验前:确定样本量和实验周期(
plan) - 实验中:有人想“看一眼结果”,判断能不能提前停
- 实验后:解读结果,判断全量、继续观察、回滚还是重做
- 审查别人写的实验结论是否站得住
- 排查分流异常、结果离谱、指标打架
典型输入:
这个实验转化率提升了 0.3 个百分点,p 值 0.04,可以全量吗?不适用场景
- 没有随机分组的前后对比、自然实验:请使用
business-root-cause-analysis,并明确说明这不是实验 - 普通漏斗分析:使用
funnel-analysis - 留存分析:使用
retention-cohort-analysis - 指标定义本身有争议:先用
metric-definition-reviewer
没有对照组、实验组和核心指标数据时,不要给实验结论。
输入信息
最少输入(汇总数据即可):
- 每组样本量
- 核心指标的结果:比例指标给成功数,均值指标给均值和标准差,比率指标(如点击率)给用户粒度的分子和分母
- 实验天数
推荐输入:
- 实验假设、分流比例、计划周期
- 护栏指标结果(错误率、性能、投诉、退款等)及可接受的恶化幅度
- 预期分流比例(非均分实验必须提供,否则无法判断 SRM)
- 用户粒度明细 CSV(可做 CUPED 方差缩减)
- 实验前同一指标的数据(CUPED 协变量)
输入字段详见 输入格式。
上下文建议
推荐使用 实验分析上下文模板。
最有价值的上下文是:分流方式、计划周期、护栏指标、实验期间是否有大促或版本发布。缺少护栏指标时,不能建议直接全量。
工作流程
- 整理输入:把用户给的数字写成
summary.json(格式见 references),或直接使用用户的 CSV。 运行脚本(能执行代码时必须运行):
python3 scripts/ab_test.py analyze --summary summary.json --html ab_report.html # 或用户粒度数据: python3 scripts/ab_test.py analyze --csv users.csv --variant-col variant --control A \ --metric paid:proportion --metric gmv:mean --guardrail refund:proportion:lower \ --cuped gmv=pre_gmv --days 14 --html ab_report.html脚本输出 JSON:
srm、comparisons、checks、verdict。verdict是规则给出的最低结论,不能比它更乐观。CSV 可以是 UTF-8 或 GBK,比例列支持 0/1、true/false、是/否、Y/N。非均分实验加
--allocation A=0.9,B=0.1;比率指标写成clicks/views:ratio。全部参数见 命令行参数。- 逐项检查有效性:对照 有效性检查清单,补充脚本无法判断的风险(污染、新奇效应、大促、版本发布、埋点变更)。
- 形成决策:在脚本结论基础上加入业务判断:收益是否覆盖成本、风险是否可控。
- 交付:Markdown 结论 + HTML 报告路径。
实验前规划:
python3 scripts/ab_test.py plan --baseline 0.07 --mde-rel 0.05 --daily-users 6000不能执行代码时:按同样的检查顺序人工分析,并在结论中写明“未经脚本计算,数值为近似”。
输出格式
## A/B 实验分析
### 1. 结论
<一句话:全量 / 逐步放量 / 继续实验 / 回滚 / 重做,以及最关键的理由>
### 2. 指标结果
| 指标 | 类型 | 对照组 | 实验组 | 变化(pp / 相对) | 95% 区间 | p 值 | 显著 |
| --- | --- | --- | --- | --- | --- | --- | --- |
### 3. 有效性检查
| 检查项 | 结果 | 说明 |
| --- | --- | --- |
| 样本比例(SRM) | | |
| 实验周期 / 整周 | | |
| 提前查看 / 提前停止 | | |
| 护栏指标 | | |
| 多重比较 | | |
| 外部干扰(大促、发版、埋点) | | |
### 4. 如何理解这个结果
<功效与最小可检测效应;不显著时说明“检测不到”和“没有效果”的区别>
### 5. 决策建议与下一步
### 6. 需要补充的数据
报告文件:<ab_report.html 路径>决策规则
脚本的 verdict.code 对应以下规则,人工判断只能更保守,不能更激进:
| code | 含义 | 默认动作 |
allocation_unknown | 没给分流比例,且实际占比明显偏离均分 | 确认设计比例后用 --allocation 重跑 |
invalid_srm | 分流异常 | 不解读任何指标,排查后重做 |
guardrail_regression | 护栏显著恶化 | 不全量,由业务方明确是否接受代价 |
primary_regression | 主指标显著下降 | 回滚 |
win_unverified | 主指标显著提升,护栏缺失 | 补护栏后再决定 |
win_needs_validation | 主指标显著提升,但周期不足、提前查看或停止、提升异常大、样本偏小,或护栏未证明在容忍度内 | 暂不放量,按 next_steps 逐项复核 |
ship | 主指标提升,护栏未见恶化,有效性检查全部通过 | 按 10% → 50% → 100% 逐步放量 |
inconclusive | 不显著 | 不下结论;按功效计算延长或重新规划 |
另外:
- 提升幅度异常大(> 30%)先怀疑数据,再庆祝
- 统计显著但收益很小:结合上线和维护成本判断
- 业务提升明显但不显著:说明需要多少样本,而不是“方向是对的,可以上”
质量标准
输出必须:
- 所有统计量来自脚本输出,或明确标注为近似值
- 比率指标的变化同时写百分点和相对变化
- 给出置信区间,不只给 p 值
- 不显著时写出当前样本的最小可检测效应,不写“新方案无效”
- 检查 SRM、周期、提前停止、护栏和多重比较
- 把“数据显示提升”和“建议上线”分开表达
verdict是win_needs_validation或allocation_unknown时,不能写成“建议放量”
完成后的下一步
- 写成报告:
data-analysis-report-writer,把--json的结果作为来源 - 向管理层汇报:
data-presentation-architect
示例 Prompt
请用 ab-test-analysis 解读这个实验,并生成 HTML 报告:
假设:新版会员页能提高开通转化率
对照组:旧版页面,10000 UV,转化 310
实验组:新版页面,10050 UV,转化 355
护栏指标:退款率、页面加载时间、客服投诉(还没汇总)
周期:7 天,原计划 14 天请用 ab-test-analysis 规划实验:
支付成功率基线 7%,希望能检测到 5% 的相对提升,每天约 6000 个下单用户进入实验,分两组。
需要多少样本、跑几天?请用 ab-test-analysis 分析 users.csv:
分组列 variant,对照组 A;主指标 paid(是否下单)和 gmv;护栏 refund(越低越好);
pre_gmv 是实验前 28 天 GMV,请用 CUPED。实验跑了 14 天。Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

