agentsclimarketplace

08 data audit

Skill lianxhcn/PXa2026a/skills/core/08-data-audit

Lecture

Install
npx -y skills add lianxhcn/PXa2026a --skill 08-data-audit

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

3 things to look at

  • 20 days oldThe repository was created 20 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
  • 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.
  • 3 stars3 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.

What its author says it does

Copied from the file, not written here

Use this skill to audit and cross-validate a data-cleaning result before trusting it, especially work produced by an AI/agent. Given the cleaning code plus before/after data, it checks the empirical red lines: is every change in sample size accounted for (a sample-selection log); is the merge match rate normal (no silent losses); were outliers blanket-winsorized instead of investigated; are variable definitions/units consistent and summary statistics in line with published papers on similar samples; and above all, were any samples dropped or thresholds picked to favor a desired result. It never modifies data; it only flags suspicious points for the researcher to review.

SKILL.md

4.3 KB, as published. Nobody here has run it

data-audit(数据审计员)

数据清洗做完之后,在你信任它、拿去跑回归之前,用它做一次体检。它专门审查一份清洗结果有没有踩「数据处理红线」——尤其针对 agent 干的活:agent 很能干,但它优化的是「完成任务」,未必是「对研究负责」。这个 skill 拿着一张红线清单逐条核对,把可疑处列出来交你复核。它不改数据,只审、只报。对应第 2 讲末尾的「不可破除的红线」。

何时用

  • agent(或你自己)清洗完一份数据,想在用它之前查一遍有没有出问题;
  • 复现别人的数据处理,想核对样本量、匹配率、变量口径对不对得上;
  • 担心清洗过程为了迎合下游结果而做了手脚(删样本、挑阈值);
  • 想给一份清洗流程配上「样本筛选日志 + 变量字典 + 一致性检查」的成品。

使用方式(提示词)

把清洗的 do / 代码,以及清洗前后的数据摘要(describesummarize、样本量)交给 agent:

这是一份数据清洗的代码和清洗前后的数据摘要:[粘贴]。请当我的审计员,逐条核对下面这些红线,把每一条的核对结果和可疑处列出来,只报告、不要改数据:

  1. 样本流失有没有账:从原始到最终,每一步样本量怎么变的?有没有哪一步无故大增大减?
  2. 合并匹配率:每次 merge 的匹配率是多少?有没有大量 master-only / using-only 没被发现?
  3. 离群处理:离群值是先查明来源再处理,还是被无差别缩尾/删除?有没有把本可修正的录入错误也一并缩尾?
  4. 变量口径与单位:构造变量的定义、单位是否前后一致?关键变量的均值/标准差,和做同类样本的已发表文献相比是否离谱?
  5. 有没有为结果做手脚:有没有在没有正当理由的情况下删掉某些样本、或专挑能凑出显著的阈值?

对每条给出:通过 / 存疑 / 不通过,以及依据。存疑处告诉我该去核对什么。

红线检查清单(skill 内建)

  • 原始数据是否只读、清洗是否全部写成代码(可复现);
  • 样本筛选日志是否完整、每步增减是否有理由;
  • 每次合并是否记录并检查了匹配率,异常是否被发现;
  • 离群是否「先查明来源、再决定」,而非无差别处理;
  • 变量口径、单位是否一致,基本统计量是否与文献可比;
  • 是否存在为迎合结果而做的可疑筛选或阈值选择。

输出约定

  • 逐条给「通过 / 存疑 / 不通过」+ 依据,不含糊;
  • 存疑与不通过项,指出该去核对什么、可能的后果,而不是直接替用户下结论或改数据;
  • 发现「样本无故大变」「匹配率异常未报警」「离群被无差别缩尾」「为结果筛样本」中任何一项,一律标为需人工复核的高风险项。

最小使用示例

用 data-audit 审一下这份清洗:原始 126 行,最终 118 行,中间只写了「drop if 缺失」。请核对样本流失有没有交代清楚、这 8 行是怎么没的。

人工检查清单

  • 审计结论只是「提示」,最终由你回到原始数据核实;
  • 高风险项(尤其「为结果筛样本」)必须人工逐条查证;
  • 审计发现的问题,回填到样本筛选日志与变量字典(配 06-repro-logger)。

来源与许可

  • 原创:本课程为「连享会 2026 暑期班·初级班」编写,Agent 无关的开放 Skill,MIT(随仓库 LICENSE-CODE)。
  • 最后验证日期:<!-- TODO: T7 彩排 Claude Code 与 Codex 双端实测后填写 -->

Keep looking

Skills are one crate of 328,083. 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.