Spreadsheets: describe, don't guess

Intermediate · 7 min · Get spreadsheet formulas with assumptions surfaced and a way to test them.

spreadsheets
verification

Ask for “a formula to total my expenses” and you get a confident answer that references columns you do not have. The model cannot see your sheet — it fills every blank with the most average spreadsheet in its training data. The discipline is in the name: describe, don’t guess — yours and the model’s.

Describe your actual structure

Before the ask, state what exists: which columns, what lives in each, which rows are data, where the summary block should sit. “Column C holds the category text, column D the amount; data rows 2 to 340.” Ten seconds of description removes the entire class of invented-column answers.

The four-part request

Structure — the real columns, ranges and one sample row.
Outcome — the exact result in sheet terms: “a running total per category for a chosen month”.
Format — how the answer must come back: assumptions first, then the formula, then a plain-language walkthrough of each part, then one test with values you already know.
Guard — “do not reference anything I did not describe; if something is unspecified, ask instead of assuming”.

The test step is not optional

A formula is a hypothesis about your sheet until proven. The request asks for a test case — “what should this show for these three rows?” — and you run it on a copy of the sheet with values you can verify by hand — a tiny dry run: three rows, five seconds, and the formula is either trusted or fixed. This is the spreadsheet version of verifying before you ship: cheap check, expensive mistake avoided.

A bad example

Ask: “Give me a formula to sum expenses by category.”

The answer assumes one header row, categories in column B and amounts in column F. Your categories are in C and your amounts in D. Wrong sheet, confident formula.

A better example

Ask: “My sheet has one receipt per row: A = date, B = vendor, C = category, D = amount; data rows 2–340. I want a per-category total for a month I pick in a cell. First state any assumptions, then the formula, then a walkthrough of each part, then one test I can run with three known rows. Do not reference columns I did not describe.”

Now the formula fits the sheet you have — and you can prove it before anyone else sees it.

Data hygiene, briefly

Describe structure, not contents. Your real employee names, salaries or customer records do not need to travel with the formula — sample rows with made-up values communicate structure just as well (redaction, in spreadsheet form), and some of it should not leave your machine at all. That principle gets its own full treatment in the safety track; here it has one form: structure in, no real data out.

Practice

Assemble a spreadsheet-help prompt

Your expense tracker holds one receipt per row — date, vendor, category and amount. You want a running total per category for a monthly report, and a formula you understand well enough to trust.

Fill every field to assemble a prompt that describes your sheet and asks for a testable formula — then compare with a strong example.

Structure — describe the sheet as it actually is

Outcome — the exact result you want

Format — how the answer must come back

Guard — what it must not do

Your prompt, assembled

Hint

What does the model not know about your sheet unless you say it?

Which part of the answer lets you prove the formula before trusting it?

Transfer

Reporting: monthly summaries from a described sheet, tested on a copy.
Cleanup: “describe the transformation, then show it applied to three sample rows”.
Handover: the walkthrough becomes the comment you leave for colleagues.

That closes the work track — all six lessons. Next up: the coding track begins with What AI can (and can’t) do with code — realistic expectations before you generate anything. Take the practice floor meanwhile for mixed repetitions.