← ClaudeAtlas

eda-quickstartlisted

Run a quick, explained exploratory data analysis on a new dataset (shape, types, missing values, distributions, target balance, correlations, leakage risks) and suggest the next modeling steps. Use when the user shares a CSV, dataframe or Kaggle or bootcamp dataset and asks to explore, understand, clean or start working on it.
simonlivson/claude-skills · ★ 0 · AI & Automation · score 60
Install: claude install-skill simonlivson/claude-skills
# EDA quickstart A first look at a dataset that is fast, reproducible and explained, so the next step is obvious. ## When to use - A new CSV or dataframe, a bootcamp exercise dataset, or "what's in this data?" before modeling. ## Inputs - The file path or dataframe, and the target column if there is one (ask if it's a modeling task). - The goal: description only, classification, regression or clustering. ## Steps 1. **Load and describe**: shape, dtypes, `head()`, memory. Note ID-like and date columns. 2. **Data quality**: missing values per column (count and %), duplicates, constant columns, obviously wrong values (negative ages, impossible dates). 3. **Univariate**: numeric summaries and histograms; top categories for categoricals; flag skew and outliers. 4. **Target**: class balance or target distribution. Say what metric that implies (for example F1 or PR-AUC when classes are imbalanced). 5. **Relationships**: correlations with the target, strongest pairs, and a few grouped means for categoricals. 6. **Leakage check**: columns that wouldn't be known at prediction time, or that encode the target. 7. **Next steps**: cleaning plan, encoding and scaling choices, a baseline model and how to split (stratified, by time, by group). ## Output format Runnable code cells (pandas, matplotlib or seaborn), each with a markdown cell above it saying what it checks and why, then a short findings list with the real numbers from the output. ## Guardrails - Report only numbers the code