data-science

Data Science / Notebook Rules

Notebook outputs are the single biggest hidden token cost in DS work — printed DataFrames, plot bytes, and stack traces all flood context.

Cell output discipline

• After running a cell that produces large output, suggest clearing it before next prompt: print only df.shape, df.head(3), or df.dtypes.
• For plots: confirm visually, then suggest plt.close() and remove the inline image from the next prompt.
• Don't paste tracebacks longer than 30 lines — paste the final exception line and the most recent frame in your code.

DataFrame work

• Use .head(n) / .sample(n) for inspection, never full prints.
• Don't print(df) on a DataFrame larger than 50 rows.
• Show schema with df.dtypes or df.info() rather than the data itself.
• For a transformation: show the chain (df.groupby(...).agg(...)) without printing intermediate results.

Code cell output

• Edit single cells in place; don't regenerate the whole notebook.
• Use # ... cell unchanged between edits.
• Don't add markdown cells unless asked.

Modeling

• Match existing framework: scikit-learn / PyTorch / TensorFlow / JAX. Don't switch.
• Don't introduce a new experiment-tracking tool (W&B, MLflow) if not present.
• Don't add hyperparameter sweeps unless asked.
• Don't print(model) for large models — show only the layer/parameter being changed.

Reproducibility

• Set random seeds only if the surrounding code does.
• Don't add pip install cells in the middle of a notebook — flag missing deps for the requirements file.

Heavy outputs to suppress

• Full model summaries (model.summary() for big networks)
• Feature importance for thousands of features (show top-20)
• Confusion matrices wider than 10×10 (show as text counts, not seaborn plot)
• Tensor pretty-prints (use .shape and .dtype only)

Never

• Don't display(df) without a row limit.
• Don't paste binary outputs (image base64, pickled bytes) into chat.
• Don't refactor a research notebook into "clean code" unless asked.
• Don't add unit tests to exploratory notebooks.