Academic Notebook

Academic Notebook Skill

Purpose

• Create rigorous, complete, academic-style Jupyter notebooks for data science and machine learning tasks.
• Produce writing that feels human and technical, not generic AI boilerplate.
• Keep all notebook content in English.

Mandatory writing style

• Never use emojis.
• Never use decorative heading separators made only of repeated hash symbols.
• Never use the em dash punctuation mark. Use commas, colons, or parentheses instead.
• Use precise, formal language and explicit technical reasoning.
• Keep markdown clean and readable, with meaningful section titles and short, coherent paragraphs.

Mandatory notebook structure

• Index section near the beginning with clickable links to major sections.
• Title.
• Introduction.
• Problem framing and objective.
• Data overview.
• EDA (Exploratory Data Analysis) with high-value visuals and concise interpretation.
• Data quality checks (missing values, outliers, duplicates, leakage risks, target imbalance where relevant).
• Preprocessing and feature engineering.
• Modeling strategy and validation design.
• Cross-validation results.
• Model comparison with a wide model set.
• Feature importance and interpretability.
• Error analysis and failure cases.
• Robustness and sensitivity checks.
• Limitations and threats to validity.
• Reproducibility notes.
• Conclusion section placeholder only (do not write the conclusion text itself).

Modeling requirements

• Always include a strong baseline first.
• Always evaluate a diverse family of models when the task permits, for example:
• Linear and generalized linear models.
• Tree-based methods.
• Ensemble methods (bagging and boosting).
• Kernel-based methods.
• Distance-based methods where appropriate.
• Neural approaches when justified by data size and objective.
• Use cross-validation as the default for model selection.
• Use stratified cross-validation for classification when class imbalance exists.
• Keep a holdout test set for final reporting whenever feasible.

Metrics requirements

• Report multiple complementary metrics, not a single score.
• Classification examples: accuracy, balanced accuracy, precision, recall, F1, ROC-AUC, PR-AUC, log loss, calibration metrics, confusion matrix.
• Regression examples: MAE, MSE, RMSE, R2, adjusted R2 (when relevant), MAPE or sMAPE when appropriate, residual diagnostics.
• Ranking or probabilistic tasks: include task-specific metrics and calibration quality.
• Discuss metric trade-offs and practical implications.

Interpretability requirements

• Always include feature importance analysis when models allow it.
• Prefer both global and local interpretability views when feasible.
• Use permutation importance as a robust model-agnostic baseline.
• Include SHAP or equivalent methods when computationally feasible.
• Validate that top features are plausible and check for leakage signals.

Advanced analysis requirements

• Hyperparameter optimization with a justified search strategy.
• Prefer advanced search tools when suitable, including GridSearchCV and HalvingRandomSearchCV.
• Bias-variance discussion grounded in observed results.
• Learning curves where useful.
• Threshold analysis for classification if decisions depend on operating points.
• Ablation studies for key feature groups or preprocessing choices.
• Stability analysis across random seeds or folds.

Library and API coverage expectations

• Use a broad scikit-learn stack when compatible with the task:
• sklearn.preprocessing
• sklearn.feature_selection
• sklearn.decomposition
• sklearn.manifold
• sklearn.pipeline
• sklearn.ensemble
• sklearn.linear_model
• sklearn.svm
• sklearn.calibration
• sklearn.neural_network
• sklearn.metrics
• sklearn.model_selection
• sklearn.experimental
• Include xgboost models when the package is available.
• Include gradient boosting models in the model portfolio.
• Favor these validation and search utilities where appropriate:
• enablehalvingsearch_cv
• StratifiedKFold
• StratifiedGroupKFold
• GroupShuffleSplit
• crossvalpredict
• traintestsplit
• GridSearchCV
• HalvingRandomSearchCV
• cross_validate
• Include dimensionality reduction visual diagnostics, at minimum PCA and t-SNE when computationally feasible.

Quality bar

• Each section must answer a clear question and end with a short interpretation.
• Plots must include titles, labeled axes, and units when relevant.
• By default, show plots inline and do not save image files to disk.
• Tables must be readable and sorted to support decisions.
• Code must be reproducible, with explicit seeds and deterministic options when available.
• Keep notebook flow logical, from problem framing to evidence-based model choice.

Deliverable constraints

• Always include both a title and an introduction near the start.
• Always show dataset snapshots before cleaning and after cleaning.
• Always serialize the final selected model as a .pkl file in a models/ directory, creating that directory if needed.
• Always include a final section named Conclusion (to be written) with placeholder prompts only.
• Do not auto-generate a final narrative conclusion in that section.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: pinthoz
• Source: pinthoz/claude-notebook-skill
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.