DataSkills

Reusable agent skills that make AI coding agents do data science correctly: explore before modelling, clean without hiding anything, and catch leakage before it fakes a great score. Each skill is a short SKILL.md plus one tested Python script. Free libraries only, no API keys.

View on GitHub ↗Example reportsMeasured results

The skills

eda-report

Profile a dataset: shape, column types, missing values, distributions, correlations and the target, saved as a markdown report with charts.

SKILL.md ↗ · example

data-cleaning

Fix what is clearly broken (spaces, "N/A" text, money strings, mixed dates, spelling variants, duplicates), flag what needs a decision, log every change.

SKILL.md ↗ · example

leakage-check

Find target leakage (columns recorded after the outcome) and train/test contamination (shared rows, groups, time overlap), with HIGH/MEDIUM/LOW findings.

SKILL.md ↗ · example

Before / after on real data

Titanic passenger list (1,309 people). Same model and pipeline, 5-fold cross-validation × 3.

ColumnsAccuracyROC AUC
Before: every column12, incl. boat, body0.9740.993
After: leakage-check first10 (removed: boat, body)0.7870.847

The 97% is fake: boat is the lifeboat a passenger got into and body the number of a recovered body. Neither is known before the ship sinks. The honest score is 78.7%.

Measured results

Problems planted on purpose into four real datasets (scikit-learn breast cancer, wine, diabetes, and Titanic without the leaking columns).

81 / 81planted leaks found by leakage-check (7 kinds)
0HIGH false alarms on 4 clean datasets and 24 correct splits
1,811 / 1,811broken cells restored by data-cleaning, 0 other cells changed
5.5 sEDA report for 1,000,000 rows × 21 columns (200k-row sample)
Planted leakFound
Copy of the target (re-coded or noisy)12/12
Value recorded after the outcome12/12
Value present for one outcome only9/9
File sorted by the target12/12
Same customer/patient in train and test12/12
Test older than train12/12
Training rows copied into test12/12

In breast cancer, 10 genuinely strong measurements are marked MEDIUM ("check when this value is known"): strong is not the same as leaking. The thresholds and planted problems were designed by me, so these numbers show the scripts do what they promise, not that they catch every leak in the wild.

Example outputs (Titanic)

EDA reportCleaning logLeakage report

Install

git clone --depth 1 https://github.com/m-shahmeer-kz/dataskills ~/dataskills && mkdir -p ~/.claude/skills && cp -r ~/dataskills/skills/* ~/.claude/skills/ && pip install -r ~/dataskills/requirements.txt

Then ask your agent things like "Explore data/sales.csv" or "Is my churn model leaking?".

MIT licence · Muhammad Shahmeer · github.com/m-shahmeer-kz/dataskills