Claude Skills by StamKavid

Records a real failure-and-fix pair as a project-local lesson, with the specifics that make it recognisable next time. Use when a bug, leakage mistake, or validation error was found and corrected and should not recur. Use when someone asks to save what was learned from an experiment that went wrong.

Separates what a model found from what would happen if a feature were changed, and catches association reported as cause. Use when someone asks whether a feature causes an outcome, or whether changing it would change the result. Use when a write-up says something drives, causes, or leads to the target. Use when a recommendation implies intervening on a variable rather than just scoring with it.

Carries a tabular machine-learning request from a plain-language ask all the way to a scored model and an honest verdict, in one turn. Use when the user says build a model, train a classifier, classify or predict or forecast a column, or detect something in a CSV, table, or spreadsheet. Use when someone wants to look at a dataset and see whether an outcome can be classified or predicted from it. Use when someone asks how well a model works, whether a result is good enough, or whether a model ...

Chooses the right chart and library for the audience and keeps it honest — no truncated axes, no misleading aggregation. Use when someone asks which chart or plot type to use, or whether to reach for matplotlib, Plotly, or Altair. Use when a figure looks misleading or a chart choice seems arbitrary.

Decides when pandas is fine and when to reach for Polars, plus vectorisation, dtype, and memory technique either way. Use when an operation such as a groupby, join, or apply is slow, or a script takes far longer than it should. Use when a dataset no longer fits in memory. Use when someone asks about Polars, vectorisation, or memory usage.

Checks whether training data still resembles the data a model will actually see, via adversarial validation and per-feature drift comparison. Use when a cross-validation score was strong but real or held-out performance fell apart. Use when someone asks whether training data still looks like production. Use when choosing a split, and again when a good score fails to transfer.

Scores a deliberately dumb baseline — majority class, mean, median, or a one-rule heuristic — so there is a number the real model has to beat. Use when someone asks whether a score is actually good, what to compare a model against, or whether a result beats guessing. Use before training or tuning any model, whenever no baseline number exists yet.

Rewrites a technical write-up as a one-page brief for a non-technical reader — no metric names, framed in money, percentages, and counts. Use when someone asks to explain results to an executive, a business audience, or anyone who does not know what AUC means. Use when a plainer version is wanted for business stakeholders. Use when a summary needs to fit on one page with the jargon stripped out.

Profiles a dataset before any modelling — schema, columns, row count, dtypes, missing values, duplicates, integrity checks, provenance, and a data dictionary. Use when someone asks what columns a file or table has, how many rows it contains, or what is inside a dataset they were just handed. Use when someone asks to profile a table or check its schema and integrity. Use when a new CSV, file, table, or spreadsheet enters a project and nothing has profiled it yet.

Stands a packaged model up as a callable endpoint with prediction logging against the live baseline, drift detection, and a rollback pointer. Hard gate — refuses full traffic without all three, and stops before any remote or cloud push. Use when someone asks to serve, deploy, or stand up a model endpoint. Use when monitoring, drift, or rollback come up for a model that is about to run live.

Measures how a model really performs — the metric at a stated operating point, a confusion matrix, calibration, and performance broken out by slice rather than one aggregate number. Use when someone asks how well a model does on held-out data, wants precision and recall at a threshold, or wants a confusion matrix. Use when checking whether a model performs worse for particular groups, segments, or subgroups.

Interprets a trained model with permutation importance and SHAP to check it learned something sensible rather than an artifact. Use when someone asks which features a trained model is actually using or relying on, what drives its predictions, or why it decided something. Use when feature importance or SHAP values are wanted for an already-fitted model. Use after evaluation and before results are shown to anyone.

Runs exploratory data analysis — distributions, relationships between each feature and the target, correlations — with a running hypothesis log. Use when someone asks for EDA, wants to explore or visualise a dataset, or wants to plot how features relate to what is being predicted. Use after a dataset has been profiled and before features are engineered.

Turns a vague data request into a precise problem — unit of analysis, target definition, the decision it feeds, and what success means. Use when someone wants to predict something but has not said exactly what, or asks what the target variable should be. Use when a model is requested with no success criterion, no metric, and no stated decision behind it. Use before touching data on a new project.

Packages a project so someone else can rerun it — pinned environment, model card, and a rerunnable artifact. Hard gate — refuses to proceed without exact pinned versions. Use when someone asks to make work reproducible, pin dependencies, or write a model card. Use when a project is being handed to another person or team.

Reads the evaluation's findings, names what is actually wrong — bias, variance, a weak slice, suspected leakage, a data problem — and routes back to the stage that fixes it. Use when a result is not good enough and someone asks what to try next. Use when deciding whether to go back and change features, change the split, or change the model, versus moving on.

Holds the shared discipline every stage of this pipeline cites — the Red Flags, the Common Rationalizations, and the Hard Gates. Use when someone pushes back on a gate, asks why a baseline is needed, or wants to peek at the test set just once. Use when deciding whether a stage should stop and ask or produce the missing work itself.

Trains, tunes, and compares candidate models against the chosen validation scheme and the baseline number. Use when someone asks to train models, tune hyperparameters, or compare algorithms such as logistic regression, random forest, XGBoost, or LightGBM. Use when picking which of several trained candidates to carry forward.

Wraps a handed-off model as a servable unit — inference contract, thin predict wrapper, reproducible Dockerfile — and proves it returns the same predictions it produced offline. Hard gate on that parity check. Use when someone asks to turn a model into a service, build a container or Docker image for it, or define its inference contract. Use before any deployment.

Cleans data and builds features with leakage prevention as the organising principle — every transform fit on training rows only, wrapped in a pipeline. Use when someone asks to clean data, encode categorical variables, scale numerics, or handle missing values. Use when engineering or transforming features ahead of training a model.

Turns evaluation and interpretation into a stakeholder-facing narrative — recommendation, assumptions, and limitations stated plainly. Use when someone asks to write up, summarise, or present results. Use when preparing findings for a product team, a manager, or any audience that will act on them.

Chooses a leakage-safe way to split data — temporal, grouped, stratified, or nested cross-validation — matched to how the rows are actually structured, and decided before any model is trained. Use when someone asks how to split into train and test, set up cross-validation, or pick between KFold, GroupKFold, StratifiedKFold, and TimeSeriesSplit. Use when repeated entities, time ordering, or nesting could make a random split leak. Use when tuning hyperparameters without contaminating the test set.

Slices residuals and errors to find where a model fails, instead of reporting one aggregate score. Use when someone asks where a model gets things wrong, or wants the worst-performing segments named. Use when an overall metric looks acceptable but the result is not yet trusted. Use when debugging why specific cases come out wrong.

Handles targets where the interesting class is rare — resampling, class weights, threshold tuning, and the metric consequences. Use when only a tiny fraction of rows are positive, or when a model never predicts the rare class at all. Use when someone mentions SMOTE, oversampling, undersampling, or class_weight. Use when accuracy looks high because almost everything belongs to one class.

Picks the evaluation metric that matches the decision and the shape of the target, and says plainly when a familiar one is wrong. Use when someone asks whether accuracy is the right thing to report, or whether to optimise for precision or recall. Use when choosing or defending a metric for classification, regression, ranking, or probability estimates. Use when a metric was picked out of habit rather than from the decision it informs.

Blends, stacks, or averages several trained models using leakage-safe out-of-fold predictions, and reports the lift over the best single component relative to fold spread. Use when someone asks to combine models, average predictions, or stack them. Use when a single model's score has plateaued and the next lever is a combination rather than a better model.

Makes a notebook safe to rerun top to bottom and safe to read out of execution order. Use when a notebook only works if cells are run in a particular order, or its current state can no longer be trusted. Use when someone asks to clean up or tidy a notebook before sharing, reviewing, or handing it over.

Finds features that encode the answer — computed from the label, from the future, or contaminated across the train/test boundary. Use when a score looks too good to be true on the first try, such as AUC near 0.99 or R-squared near 1. Use when one feature dominates the importance ranking. Use when a feature was built from aggregates, neighbours, or anything derived from the thing being predicted.

Reports every score with its spread across folds or seeds, and says whether a gap between two numbers is bigger than that spread. Use when someone asks whether an improvement is real or just noise, or how confident to be that one model beats another. Use when two results are being called different or the same without a spread behind the claim.