how to analyze a spreadsheet with claude, step by step.

elisabeth hitz · july 14, 2026 · 5 min read

to analyze a spreadsheet with claude accurately, get the data clean, tell it what each column is, and ask one focused question at a time instead of dumping a messy workbook and asking something vague. the workflow matters more than the model. follow the steps in order and the fumbling stops.

most people do the opposite. they paste the whole file, ask "what does this tell me," and get a confident summary that is partly invented. here is the sequence that gets you a real answer instead.

step 1: get the data clean

before you ask anything, hand over data the model can actually read. that means one table, not a workbook. export the sheet you care about as csv, or copy the range and paste it in. delete the merged title rows, the blank spacer rows, and any total or subtotal rows sitting inside the data.

this one step prevents most errors. a clean csv reads straight; a designed excel file gets misread. that whole tradeoff is the subject of csv vs excel for ai, and it is the foundation everything else here rests on.

step 2: tell claude the shape of the data

the model cannot see your intent, only the text. so tell it what it is looking at in one line before you ask your question. name the columns and say what each one holds:

"this csv is my monthly sales. col a is the month, col b is units sold, col c is revenue in usd, col d is the rep name. one row per month per rep."

that sentence is doing real work. it stops the model guessing which column is revenue and which is units, which is exactly where wrong numbers come from. skipping it is one of the quiet reasons your ai is not working.

step 3: ask one focused question at a time

do not ask for everything at once. a single question gets a clean, checkable answer. a pile of questions gets a blur. work through them in order:

• start with a summary. "give me the totals: total revenue, total units, and the number of rows you counted." the row count tells you it read the whole table.
• then ask for the trend. "is revenue going up or down month over month? give me the number for each month."
• then hunt anomalies. "which rows look off, a blank cell, a zero where you would not expect one, or a value far from the others? list the row numbers."

each answer builds on the last, and each one is small enough to verify before you move on.

step 4: ask it to show its working

always make the math visible. instead of "what is total revenue," ask "what is total revenue, and show the monthly figures you added to get there." now you can see the numbers it used, not just the result. if a monthly figure is wrong, you catch it immediately instead of trusting a total that came from bad addition.

this also flushes out silent misreads. if claude added a stray total row you missed, the working will show a figure that does not belong, and you will know to go clean the file again.

step 5: verify a couple of figures yourself

pick two numbers and check them by hand. add up the first few rows yourself and compare. glance at a total you already trust. it takes thirty seconds and it is the difference between using the analysis and hoping it is right. if both spot checks match, the read was clean and you can trust the rest. if one is off, the file still has a problem and you fix the input, not the prompt.

the common failure, in one line

almost every bad spreadsheet analysis comes from the same move: pasting a whole messy workbook and asking a vague question. the model gets tangled in merged cells and multiple tabs, guesses at what you meant, and hands you a fluent answer built on a misread. clean file, named columns, one question, shown working, verified numbers. that is the entire method. the rest of setting up data cleanly lives in why the file format you feed claude decides the answer and in how to use claude with spreadsheets without the errors.

common questions

can claude spot errors in my data, or just summarize it?

it can flag likely errors well: blanks, zeros, values far outside the normal range, dates that do not parse. ask it directly to list the row numbers that look off, then check those rows yourself. it points you at the suspects; you confirm.

how do i get it to find a trend over time?

make sure your date or month column is clean and ask for the figure per period, not just "is it going up." seeing the number for each month lets you confirm the trend is real instead of taking its word for the direction.

why does it give a different answer when i ask the same thing twice?

usually because the question was vague enough to leave room for interpretation, or the file was messy enough to be read two ways. tighten the question and clean the file, and the answer stabilizes. asking it to show its working also makes the difference easy to spot.