claude-data-analyst

Scan a dataset for significant anomalies — outliers, distribution shifts, impossible values, and unusual groupings. Use when the user wants a first-pass integrity and anomaly sweep of a CSV/Parquet/Excel file before deeper analysis.

Detect and compute correlations between numeric variables in a dataset. Use when the user wants to see how variables in a CSV/Parquet/Excel file move together — Pearson, Spearman, or Kendall — with a short report flagging the strongest positive and negative pairs.

Generate a data dictionary for a dataset, combining automatic profiling with the user's description of what the data represents. Use when the user wants documentation of columns — names, types, semantic meaning, units, allowed values, and nullability — for a CSV/Parquet/Excel file.

Identify what the user is trying to analyse, diagnose gaps in the current dataset, propose external data sources that could fill them, then plan and implement the enrichment. Use when the dataset alone can't answer the user's question and extra context (reference data, lookups, joinable public datasets) is needed.

Produce a parametric PDF report describing a dataset — size, schema, distributions, key statistics, and findings from other skills — compiled via Typst. Use when the user wants a shareable, print-ready document about their data, not a one-off markdown summary.

Scan a dataset for signs that it has been pre-cleaned, normalised, imputed, smoothed, deduplicated, or otherwise processed before the user received it — data that is "suspiciously clean". Flag findings so the user knows whether they're analysing raw reality or someone else's editorial choices.

Take a user-stated hypothesis and test it against the data, producing a report stating whether the data supports, refutes, or is inconclusive about the claim. Use when the user has a specific question or claim they want to interrogate against a dataset.

Test relationships among three or more variables simultaneously — partial correlations, controlled effects, multicollinearity, interaction terms, and dimensionality reduction. Use when a pairwise correlation sweep isn't enough and the user wants to know how variables behave together, which effects survive when others are held constant, and which clusters of variables move as one.

Scan a dataset and flag columns or values that appear to contain personally identifiable information (PII). Use when the user wants a quick privacy audit of a CSV/Parquet/Excel file before sharing, publishing, or ingesting into another system.

Describe and assess the sample size of a dataset — not just row count, but effective sample size per question the user wants to answer. Flags underpowered segments, imbalanced classes, small-n group cells, and gives a concrete "you can / cannot reliably claim X from this data" verdict.

Set up a "talk to your data" workspace in the current repo — discover local data files, load them into a DuckDB database, and append a CLAUDE.md block telling future Claude sessions how to query it. Use when the user wants to make a repo's data conversationally queryable without wiring up a full BI stack.

Compute and interpret standard deviation (and related spread measures — variance, IQR, MAD, CV) for numeric columns in a dataset. Handles sample vs. population formulas, grouped/stratified computation, and flags columns where SD is misleading (heavy skew, outliers, near-constant values).

Identify and report the major trends a dataset depicts — directional changes over time, growth rates, seasonal patterns, segment shifts, and emerging categories. Use when the user wants the headline "what is this data saying" narrative rather than a specific test.

Scan one or more datasets for data-type inconsistencies that would block analysis or relational/graph database loading — mixed types within a column, the same logical field typed differently across files, string-encoded numbers/dates, inconsistent null sentinels. Report findings, and either delegate the fix to a Claude-Data-Wrangler skill or apply small wrangling in place.