Data Scientist Skills — Data skill for Claude Code

Data Scientist Skill Set

Skills for Claude Code that take a tabular or time-series dataset plus a business objective through to a validated model, without skipping the steps that usually get skipped.

Scope: tabular data and time series. NLP and computer vision are out of scope.

Why these exist

Most "the model never reached the target" stories trace back to omissions, not to a missing algorithm: a leaky column nobody questioned, imputation fitted before the split, a random split on time-ordered data, no baseline, model comparison on a single split, accuracy on an imbalanced target, tuning against the test set. The orchestrator turns each of those into a gate that must produce evidence before the next stage starts.

Layout

Skill	Stage it owns	What it holds
data-scientist	Orchestrator	Problem framing, stage order, which step skill to call when, the gate at each stage, the decision log, the final test-once rule, hand-off and report templates
ds-data-audit	Stages 1 and 3	A: structure audit and stateless cleaning (types, sentinels, spellings, exact duplicates, impossible values). B: leakage hunt with record-timing verdicts and relationship EDA, on training data only
ds-validation-design	Stage 2	Split strategies by data structure, CV schemes, carving and freezing the test set, noise floor
ds-missing-values	Stage 4	Mechanism diagnosis, the full imputation menu with a decision table
ds-outliers	Stage 5	Detection menu, the error / extreme / signal decision, treatment menu
ds-feature-engineering	Stage 6	Encoding, transforms, scaling, date and aggregate features, automated generation, selection
ds-model-selection	Stage 7	Baselines, candidate families by problem type, imbalance handling, learnability verdict
ds-method-comparison	Stage 8 + screening at 4-7	Local screening with two proxy models, joint conditional pipeline search, selection rules that resist noise
`d