The human-formatted spreadsheet problem

1 · Learn the move · Context + audience

The most common data file in science is not an instrument file. It is a spreadsheet a human formatted for another human to look at: merged title cells, color as meaning, units tucked into a header sentence, bold rows that are subtotals, and a Notes column holding the only record of which samples were dropped. A model handed this cold assumes tidy data - one header row, one observation per row - and reads garbage with total confidence. The fix is context before parsing: tell the model what the sheet is, who formatted it and for whom, and that nothing about its structure may be assumed. Then ask for a structure report - not an analysis - as the first output. Thirty seconds of context converts the model from a confident misreader into a careful surveyor, which is the only useful thing it can be on a sheet like this.

This spreadsheet was formatted by [person/role] for [human audience], not for parsing. Known hazards: [merged cells / color coding / subtotal rows / notes column / whatever you know]. Do not analyze anything yet. First produce a structure report: how many header rows and where the real one is, what each column actually contains, which rows are not data (titles, subtotals, notes), where units live, and anything whose meaning you cannot determine - marked UNKNOWN, not guessed. Paste: [sheet as text]