Spreadsheet triage skill

Description

A house entry. Written by the people who run Promptea, published here so the catalogue has something to show and so the required fields have a worked example.

The failure mode with agents and spreadsheets is not bad cleaning, it is confident cleaning. The model silently drops eleven rows it could not parse and reports success. This skill puts a census before the cleanup and makes the losses visible.

The instruction body

Before changing a single cell, produce a census:

  rows total / rows with a complete key
  columns, and for each: inferred type, null count,
    distinct count, and the three most common values
  merged cells, by range
  rows above the real header
  columns that are entirely empty
  dates stored as text, and the formats found

Then propose a cleanup plan and STOP. Do not execute it.

When the plan is approved, execute it and emit a second census
in the same format, plus a diff of the two. Every row that
disappeared must appear in a "dropped" list with the reason.
A cleanup that cannot account for its losses is rejected.

Never overwrite the source file. Write to a new file and keep
the original untouched.

Limits

The census on a very wide sheet is long enough to crowd the context window. Cap it at the first sixty columns and say that you did. Needs pandas or an equivalent in the runtime.