dataskills
Agent skills that make AI coding agents do data science correctly: EDA reports, logged data cleaning and leakage checks, with tested Python helpers
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/m-shahmeer-kz--dataskills)HTML
<a href="/listings/m-shahmeer-kz--dataskills"><img src="/api/listings/m-shahmeer-kz--dataskills/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
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
That copies the three skills to `~/.claude/skills/` (available in all your projects). For one project only, copy them to `.claude/skills/` inside that project instead. Other agents: put the `skills/` folders wherever your agent reads `SKILL.md` files, or tell it to read `skills/<name>/SKILL.md`.
Then just ask, for example: *"Explore data/sales.csv"*, *"Clean this file"*, *"Is my churn model leaking?"*. The agent picks the skill from its description. You can also call one directly: `/leakage-check`.
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
README.md · README · 125 lines
- # DataSkills
- [](https://github.com/m-shahmeer-kz/dataskills/actions/workflows/tests.yml)
- [](LICENSE)
- **[See the example reports ↗](https://m-shahmeer-kz.github.io/dataskills/)**
- Reusable **agent skills** that make AI coding agents (Claude Code and other tools that read `SKILL.md`) do data science work **correctly**: explore before modelling, clean without hiding anything, and catch leakage before it fakes a great score.
- | Skill | Use it when | What you get |
- |---|---|---|
- | [`eda-report`](skills/eda-report/SKILL.md) | a new dataset arrives | a markdown report with charts: shape, column types, missing values, distributions, correlations, target |
- | [`data-cleaning`](skills/data-cleaning/SKILL.md) | the data has format problems | a cleaned file + a log of **every** change (before → after) and a list of things to check |
- | [`leakage-check`](skills/leakage-check/SKILL.md) | before trusting any model | HIGH / MEDIUM / LOW findings: target leakage and train/test contamination |
- Each skill = short instructions (`SKILL.md`) + one tested Python script. Free libraries only (pandas, numpy, matplotlib, scikit-learn). No API keys.
- ## Install (one line)
- ```bash
- 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
- ```
- That copies the three skills to `~/.claude/skills/` (available in all your projects). For one project only, copy them to `.claude/skills/` inside that project instead. Other agents: put the `skills/` folders wherever your agent reads `SKILL.md` files, or tell it to read `skills/<name>/SKILL.md`.
- Then just ask, for example: *"Explore data/sales.csv"*, *"Clean this file"*, *"Is my churn model leaking?"*. The agent picks the skill from its description. You can also call one directly: `/leakage-check`.
- ## Why this exists
- AI agents write pandas fast, but they repeat the same quiet mistakes that a senior data scientist would catch:
- - training on a column that is only filled in **after** the outcome, and celebrating a 97% score;
- - `fillna(df.mean())` or `StandardScaler().fit(X)` **before** the train/test split, so test data leaks into training;
- - a random split on time data, or the same customer in train and test;
- - `dropna()` that silently deletes most of the rows, or "cleaning" that nobody can check afterwards.
- A skill gives the agent the right checklist **and** a tested tool at the moment it needs them. Instructions are short on purpose: what to do, what comes out, what not to do.
- ## Before / after (real data)
- The Titanic passenger list ([`examples/titanic3.csv`](examples/titanic3.csv), 1,309 people). Same model, same pipeline, 5-fold cross-validation × 3 ([`examples/titanic_before_after.py`](examples/titanic_before_after.py)):
- | | Columns used | Accuracy | ROC AUC |
- |---|---|---|---|
- | **Before**: agent uses every column | 12, including `boat` and `body` | **0.974** 🤩 | 0.993 |
- | **After**: `leakage-check` first | 10 (`boat`, `body` removed) | **0.787** | 0.847 |
- 97% looks great and is **fake**: `boat` is the lifeboat a passenger got into, and `body` is the number given to a recovered body. Neither is known before the ship sinks. The skill flagged both as HIGH, with plain-language reasons:
- ```text
- HIGH missing_predicts_target boat Just knowing whether `boat` is empty predicts the target (AUC 0.971).
- Values that get filled in only after the outcome look exactly like this.
- HIGH one_sided_value body When `body` is filled in, the target is `0` in 100% of 121 rows
- (overall only 62%).
- ```
- The honest 79% is the number you can actually use. Full outputs of all three skills on this file: [`examples/output/`](examples/output/) ([EDA report](examples/output/eda/eda_report.md) · [cleaning log](examples/output/cleaning/cleaning_log.md) · [leakage report](examples/output/leakage/leakage_report.md)).
- ## Measured results
- Problems were planted on purpose into four real datasets (scikit-learn's breast cancer, wine and diabetes, plus Titanic without `boat`/`body`) and the scripts had to find or fix them ([`benchmarks/run_benchmarks.py`](benchmarks/run_benchmarks.py), [`results.json`](benchmarks/results.json)).
- **leakage-check**: **81 of 81** planted leaks found, across 7 kinds: copy of the target, value recorded after the outcome, value present for one outcome only, file sorted by target, same group in train and test, test older than train, rows copied into test. **0 HIGH false alarms** on the 4 clean datasets and on 24 correct splits (group and time). Notes:
- - In breast cancer, 10 genuinely strong measurements are marked MEDIUM ("check when this value is known"). That is the intended behaviour: strong is not the same as leaking.
- - The wine file really is sorted by class, and the check says so.
- **data-cleaning**: **1,811 of 1,811** planted broken cells restored (spaces, "N/A"-style text, `$1,234.50` money strings, `-999` placeholders, `MALE`/`male ` spelling variants, mixed date formats), **74 of 74** duplicate rows removed, **0** untouched cells changed. On the clean originals it changed nothing, except 2 real problems it found in Titanic (4 names with trailing spaces; `Sweden Worcester, MA` with a double space).
- **eda-report**: Titanic in 1.8 s; a 1,000,000-row × 21-column table in 5.5 s (profiled on a 200,000-row sample).
- **Tests**: 32 pytest tests, including a check that every `SKILL.md` follows the [Agent Skills](https://agentskills.io) format, run on GitHub Actions.
- **Limits, honestly**: the planted problems and the thresholds were designed by me, so these numbers show the scripts do what they promise, not that they find every leak in the wild. Leakage that looks like a normal feature (a strong, plausible column recorded later) can only be found by asking *"would I know this value at prediction time?"*, which is why the skill makes the agent ask it. I did not measure how much agents improve with the skills (that needs paid model APIs).
- ## Repository structure
- ```text
- dataskills/
- ├── skills/
- │ ├── eda-report/
- │ │ ├── SKILL.md when to use · steps · outputs · mistakes to avoid
- │ │ └── scripts/eda_report.py profile + markdown report + charts
- │ ├── data-cleaning/
- │ │ ├── SKILL.md
- │ │ └── scripts/clean_data.py safe fixes + change log
- │ └── leakage-check/
- │ ├── SKILL.md
- │ └── scripts/leakage_check.py target leakage + train/test contamination
- ├── tests/ pytest for each script + SKILL.md format check
- ├── benchmarks/run_benchmarks.py planted-problem benchmarks -> results.json
- ├── examples/ titanic3.csv, before/after script, example outputs
- └── docs/ example reports page (GitHub Pages)
- ```
- ## How a skill works (in one minute)
- 1. At start-up the agent reads only each skill's `name` and `description` (a couple of lines).
- 2. When your request matches a description ("clean this CSV"), the agent loads that skill's `SKILL.md`.
- 3. It follows the steps, runs the script in `scripts/` (the script does the heavy, tested work), reads the report it produced, and explains the result to you.
- ## Use the scripts without an agent
- ```bash
- python skills/eda-report/scripts/eda_report.py data.csv --target y --out reports/eda
- python skills/data-cleaning/scripts/clean_data.py data.csv --out data/clean.csv --log-dir reports/cleaning --dry-run
- python skills/leakage-check/scripts/leakage_check.py data.csv --target y --out reports/leakage # exit 2 if HIGH
- ```
- ## Develop
- ```bash
- pip install -r requirements.txt pytest
- python -m pytest -q # 32 tests
- python benchmarks/run_benchmarks.py # ~35 s
- python examples/titanic_before_after.py
- ```
- New skill: add `skills/<name>/SKILL.md` (frontmatter `name` = folder name, a `description` that says *Use when...*, and the four sections) plus one script in `skills/<name>/scripts/` with tests. `tests/test_skill_format.py` checks the format.
- ## Data
- `examples/titanic3.csv`: the Titanic passenger list compiled by Thomas Cason from the *Encyclopedia Titanica*, distributed by the Vanderbilt University Department of Biostatistics (Frank Harrell). Copy taken from [amueller/scipy-2016-sklearn](https://github.com/amueller/scipy-2016-sklearn) (identical to a second independent copy). Benchmark datasets ship with scikit-learn.
- ## License
- MIT © 2026 Muhammad Shahmeer
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.