Spreadsheet Whisperer

People stare at spreadsheets and see rows. They paste data into an LLM and get a narration of what they just pasted. Neither is useful. The gap between "here's data" and "here's what this data means for your decisions" is what this skill fills, by running a five-step pipeline — health check, profile, SCAN, visualise, recommend — instead of summarising what's already on screen.

Data health comes first because bad data produces bad insights. Missing values get reported by column and percentage. Duplicates (exact and near-duplicate same-entity-different-formatting) get flagged. Format inconsistencies get standardised silently and reported — currency symbols stripped, dates ISO-normalised, common null markers (N/A, blank, -) unified. Outliers get flagged but not auto-fixed because removing them is a judgement call you should make, not the skill. Datasets under 20 rows switch to descriptive mode — no correlations, no distributions, no false aggregates from too little data.

The SCAN framework finds the signal. Segments — are there meaningful groups, top vs bottom, regions, time periods, product categories, and what differs between them. Changes — what's moving over time, growth rates, acceleration, inflection points. Anomalies — what doesn't fit the pattern, with each one flagged for investigation rather than smoothed away. Numbers that matter — the three to five KPIs someone would actually pin to a dashboard from this data, calculated and compared to benchmarks where context allows. The output prioritises what's surprising, not what's exhaustive.

Data type classification changes which analyses are valid. Time-series enables trend detection, seasonality decomposition, and forecasting. Cross-sectional enables comparison, segmentation, and ranking. Panel data enables both. The skill classifies the data first, runs the analyses the classification supports, and refuses to run the ones it doesn't (no forecasting from a single snapshot, no segment comparison when the data only has one entity). This catches the most common analysis mistake — applying the wrong technique to the wrong data shape.

You get a health check with auto-fixes named, a profile that highlights what's interesting (not every stat), the four-lens SCAN findings, two or three specific chart recommendations chosen to fit the data type rather than the generic line-chart-of-everything default, and three to five recommendations with the format Recommendation/Based on/Expected impact/Priority. The skill leads with the surprise, quantifies every claim, and ends with one question back — "to go deeper, I'd want to see [specific additional data]" — because the next iteration of an analysis usually matters more than the first. If the data is too small or noisy to draw conclusions, the skill says so. A confident wrong insight is worse than an honest uncertainty.