Analyze spreadsheets with code checked math

What it does

Chat assistants are bad at arithmetic. Ask one to total a column of invoices and it will produce a confident figure that may be off by hundreds. This skill removes that risk. It takes over the job of reading your file, running the calculations in code, and returning results you can audit. The model handles the reading and the writing. Code handles the counting, summing, averaging and comparing.

It suits sales reports, budget sheets, survey exports, gradebooks, bank statements and any messy CSV you need to make sense of. Use it when the numbers matter and when a wrong total would cost you. Skip it for quick prose tasks where no math is involved.

What you feed it

Give it three things:

1. The file itself, as xlsx, csv or a Google Sheets export.
2. The question you want answered, in plain words. For example: which region grew fastest last quarter, or how many rows have a missing email.
3. Any rules about the data, such as which column is the date, which currency you use, or whether the first row is a header.

If the file is dirty, say so. Duplicate rows, merged cells and stray totals in the middle of a sheet are common and the skill needs to know they are there.

How to run it

1. Install the skill in your agent environment and confirm it loads. Point it at the folder where you keep data files.
2. Open a session and attach your spreadsheet. State the question and any column rules in the same message.
3. Let the skill inspect the structure first. It should list sheets, column names, row counts and data types before answering anything.
4. Ask for the cleaning pass. It flags duplicates, blanks, mixed formats and text masquerading as numbers, then shows you what it plans to change.
5. Approve the cleaning, then request the analysis. Every figure comes back with the code that produced it.
6. Spot check. Pick two or three numbers and verify them by hand or with a pivot table.
7. Ask for the write-up: a short summary, a table of results, and a chart if you need one.

What a good result looks like

You get a plain answer at the top, then a table of the numbers with the code that generated them underneath. Cleaning decisions are listed, not hidden. If a column had 14 blank cells, you see that. If two rows looked like duplicates, you see which ones and why.

A good run also tells you what it could not do. If the dates are stored as free text in four different formats, it says so and asks how to treat them rather than picking one silently. You should be able to hand the output to a colleague and have them reproduce every total from the file. That is the whole point: numbers you can defend, not numbers that sound right.