I inherit spreadsheets from other teams constantly and I always burn an hour eyeballing them for landmines — mixed date formats, a 'Total' row hiding in the data, a column that's 30% blank — before I dare build anything on top.
Build a prompt that takes a {{csv_sample}} (header + a chunk of rows) and an optional {{intended_use}} and returns a data-quality report: a one-line description per column with inferred type, the likely primary key, suspected issues (nulls, duplicates, inconsistent formats, outliers, summary rows mixed into data), and 3-5 concrete cleaning steps in priority order.
Hard rules: it must reason only from the rows actually shown, never assert a statistic it can't compute from the sample, and clearly separate 'observed' from 'suspected'. When the sample is too small to judge, it says so. The strongest solution catches the subtle landmine, not just the obvious nulls. Link it and run it on a deliberately dirty sample.