菜单
Runs an honest EDA and signal audit — distributions first, heavy-tail handling, verdict-style mini-tests (CONFIRMED/OPPOSITE/MIXED/FALSE), sample-size floors. Use before modeling, when testing whether a popular belief shows up in data, or when a correlation looks too good.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Finds label leakage and designs honest validation — leakage taxonomy, grouped and time-aware splits, base rates, and the attack-your-own-model checklist. Use before trusting any metric, when a score looks too good, or when features and labels share time windows or origins.
The FlyRank internship datasets — the 30k-row starter CSV and its gotchas, the ~79M-row warehouse release tables and grains, panel warnings, access, and iteration rules. Load for EVERY task that touches the data. (Project-specific: delete this folder when reusing the skill library elsewhere.)
Works with datasets far too big to download or load in pandas — SQL over remote Parquet with DuckDB, aggregate-then-model, iterate on samples. Use when a dataset has millions of rows, lives on a remote host (hf:// or s3), or a notebook runs out of memory or hits rate limits.
Builds the transparent rule-based baseline every model must beat — a hand-written score with reason codes, ranked output, and precision@K evaluation. Use before training any model, or when someone reports model results with nothing to compare against.
Deploys a static page (research paper, portfolio piece) for free from a GitHub repo using GitHub Pages — setup, file layout, verification, and recording the final URL. Use when publishing a finished write-up as a public web page.
Teaches the working method for using an AI coding assistant on analysis tasks — one task per fresh conversation, search before assuming, interview before big plans, validate everything, commit per finished task. Use at the start of any assignment, whenever a conversation gets long and confused, or when the assistant starts guessing.
| name | auditing-signals |
| description | Runs an honest EDA and signal audit — distributions first, heavy-tail handling, verdict-style mini-tests (CONFIRMED/OPPOSITE/MIXED/FALSE), sample-size floors. Use before modeling, when testing whether a popular belief shows up in data, or when a correlation looks too good. |
Look before deciding. A signal audit asks one question per test: "does the data actually show the story people tell?" — and accepts NO as a valid, publishable answer.
1. Distributions first. Plot or describe every field you'll test. Web/traffic metrics are almost always heavy-tailed: a few giants, a long tail of tiny values. That one fact changes everything below.
2. Handle heavy tails before correlating. Plain (Pearson) correlation on raw heavy-tailed values is dominated by the giants and can flip sign after a log transform. Default to:
log1p() the traffic-like columns, or3. One mini-test per signal, with a verdict. Structure every test identically:
4. Respect sample-size floors. No verdict from a bucket with fewer than ~50 rows (~30 for cross-cuts). A huge ratio from a tiny cell is noise wearing a costume — say "insufficient data" instead, that's a finding too.