Spreadsheet Answers You Can Check

Ask an AI chat what your sales file says and you'll get an answer in seconds. It will sound sure of itself. And some of those numbers will be wrong, in ways you can't see unless you redo the math yourself.

This tutorial shows you why that happens, how to catch it, and how to use The Spreadsheet Analyst, a plugin for Claude that calculates every figure with code against your real rows and tells you up front whether your file can answer the question at all. Everything shown below is real output from the plugin's demo files. Not on Claude? Sections 02, 03, and 10 work on their own.

The Spreadsheet Analyst at a glance

The Spreadsheet Analyst reads an Excel file, a CSV, or rows you paste into the chat and tells you, in plain language, what the data actually says. Before it answers, it gives you a verdict on whether the data can support your question. Then it does the job you asked for: a quick read, a direct answer, a comparison, a cleanup, a report you can send, or research statistics with an APA write-up.

What you walk away with

• A one-screen read of any spreadsheet: the headline, what's driving it, what looks odd, and what the file can't tell you.
• A verdict before every answer, so you know when a number is safe to repeat.
• Numbers that match your sheet, because code calculated them from its rows.
• Finished files when you need them: a printable report with charts, an Excel workbook with live formulas, or a cleaned copy with a change log, each with a short covering note.
• For school and research: the right statistical test, an APA-style sentence, a results paragraph, and a table.

What it does well

Where it falls short

• Exact math needs a setup that can run code. Without one, it sticks to small tables and labels every figure as hand-calculated.
• It reads the values your spreadsheet last saved and won't recalculate formulas.
• Older formats (.xls, .numbers, .ods) and tables inside PDFs need re-saving as .xlsx or .csv first. A live Google Sheet needs downloading unless you have a drive connector.
• It won't claim cause and effect from data that only shows things moving together, and it won't invent benchmarks like "the industry average is 2%".
• Advanced research designs (mixed or logistic models) get a named method and template, not a computed result.
• The portable prompt for other AIs carries the method but not the scripts, so its math is only as good as that AI's own code tool.

Who it fits

Business owners with sales exports, budgets, or ad reports. Employees handed a file and a deadline. Students and researchers who need the right test and a write-up they can defend. People tracking their own money. If you've ever pasted a table into a chatbot and wondered whether the total was right, it's for you.

Why AI gets spreadsheet numbers wrong

Language models predict text. When one adds up a column "in its head", it's reaching for a plausible number, and it's often close. Close is the dangerous part: a total that's 3% off looks exactly like one that's right. A typical "paste your rows" prompt also never checks for title rows, subtotals, or duplicates, or asks whether the data can answer the question at all.

How to spot a wrong number

• No basis given. "Revenue was $184,000" with no mention of which rows or column. A trustworthy figure says what it's a sum of.
• It doesn't match your sheet's TOTAL row, and nobody explains why. A good answer names the gap: duplicates, numbers stored as text, a stale total.
• Suspiciously round numbers. Real data produces figures like $30,257.33. "About $30,000" is a guess.
• Percentages of tiny groups. "60% of Team East is satisfied" means little if Team East is five people.
• Benchmarks you never supplied, and cause-and-effect words like "drove" or "led to". Your file can show two things moving together. It usually can't show why.

What "messy" looks like in a real export

Files from point-of-sale systems, accounting software, survey tools, and bank portals each add their own clutter. These are the problems The Spreadsheet Analyst looks for, and what each one does to your numbers if nobody notices.

Some fixes happen automatically because they don't change meaning, like skipping a title line or reading "$129.00" as 129. Others need your OK, because only you know the answer: are those two identical rows a double entry, or did the same customer really order twice? Here's the cleanup plan it produced for the demo sales file:

The verdict comes before the answer

This is the step a typical "summarize my data" prompt skips, and it's the one that saves you from the awkward meeting. Before any analysis, The Spreadsheet Analyst tells you in one to three lines whether your data can support what you asked.

"Fix first" always puts a size on the damage, because "overstated by 1.4%" gets fixed where "some quality issues" gets ignored. "Can't answer from this" names the column that would fix the problem, then answers the closest question it can. Disagree? Say so, and it goes with your call.

Push it to "prove the campaign worked" and you'll get the before-and-after number, a sentence that's safe to paste into an update, and what data would settle it next time. It won't hunt through filters for a flattering cut.

Six modes, no special wording

You don't pick a mode from a menu. Ask the way you'd ask a colleague and it routes itself.

It also adapts to the kind of data. A gradebook gets asked whether blank scores count as zero. An ad report gets click rates computed from totals, with a warning when that ranks channels differently from a dashboard's weekly averages. You don't need to know those traps exist. An optional four-question setup ("set up my spreadsheet analyst") saves your date format and preferences so it never asks again.

Walkthrough: the demo file, with real numbers

The plugin ships with demo-store-sales.xlsx: three months of sales from a small coffee equipment shop. Type "show me a demo" and it runs a Quick Read on it. Before saying a word, it read the file properly and found a lot of clutter for 305 rows:

• Row 1 is a title line; the real headers are in row 2.
• Row 305 is a total row. Its numbers equal the sum of everything above it, so it's left out.
• Five Revenue cells hold numbers stored as text, like "$129.00". Excel's own SUM would skip them.
• Rows 39 and 127 are exact copies of the rows above them, order number included.
• No sales at all on 3 May or 8 June. Closed, or missing rows? Worth checking.

The duplicates sit inside the data, so any total that includes them is too high. That makes the verdict Fix first, with the damage measured: $86.99. Here's the read it produced:

Read it like an analyst would

The flattering fact is that revenue grew every month. The useful fact is that orders didn't: 104 in May, 103 in June. Growth came from bigger baskets, and the bigger baskets came from one $449 espresso machine. Then look at "Worth a closer look": a single order for three machines on 24 June. Without it, June is up 8.6% on May, not 22.7%. That's the kind of finding that changes a plan, and it's why the plugin ranks a risk above a nice-looking total.

Notice what it left out: no "trend" from three months of data, no invented industry average. And it shows the typical order ($54.99) beside the average ($100.86), because a few machines drag the average far above what a normal customer spends.

Reports, workbooks, and the covering note

Chat answers are fine for you. The moment someone else needs the numbers, you want a file. Ask for "a report for my boss" or "an Excel file I can keep updating" and it picks the format from what you said.

The HTML report

Best for sending, printing, or saving as a PDF. The headline states a finding with a number, and every chart has a takeaway subtitle and a source note naming the rows it summarizes. Bars start at zero, always.

The Excel workbook

Best when you'll keep working with the numbers, or someone will ask where they came from. Summary figures are live formulas pointing at a Data sheet, so clicking Total revenue shows =SUM(Data!$H$2:$H$301). Before handing it over, the plugin recalculates every formula from the saved file and checks it against its own answer. If Claude's environment lacks the Excel chart library, the workbook says so and the charts go in the HTML report.

The covering note

A file on its own is half a deliverable, so every report or workbook comes with a short note you can paste into an email. For the demo file, sent to a manager:

Everything it creates lands in spreadsheet-analyst/outputs inside the folder you're working in, with plain names like demo-store-sales - report - 2026-09-25.html.

Academic and research mode

Students and researchers need to know whether a difference is real, and how to write it up so it survives a thesis defense. This mode asks at most three questions about your design, then picks the test and tells you why. Here it is on the demo survey, asked whether likelihood to recommend differs by plan:

In plain words: the plans really do score differently, and plan accounts for about a third of the spread in scores (η² = .33). With only 18 Team users, the verdict carries a caveat for your limitations section.

Where it draws the line

• It won't hunt for significance. Dropping outliers until p falls under .05, or trying tests until one works, is p-hacking, and it declines.
• It never alters or drops data to change a result. A 999 on a 0 to 100 scale is an entry error and gets treated as missing; an unusual but real score stays in.
• It reports non-significant results properly: "no evidence of a difference", never "no difference".
• It leaves the interpretation to you and reminds you to follow your institution's policy on AI assistance.

Installing it in Cowork, claude.ai, and Claude Code

Pick the route that matches where you use Claude. Anthropic has changed its menus several times this year, so if a label doesn't match your screen exactly, look for the closest equivalent.

Claude desktop app (Cowork)

Plugins in Cowork need a paid Claude plan (Pro, Max, Team, or Enterprise). Open Customize in the sidebar, then Plugins, choose the upload option, and select the-spreadsheet-analyst.plugin. If it says only .zip files are accepted, rename the file to the-spreadsheet-analyst.zip and upload again; a .plugin file is already a zip, so nothing breaks. Team and Enterprise admins can add it for everyone under Organization settings, then Plugins & skills, then Add (that route asks for a .zip under 50 MB).

Claude on the web (claude.ai)

Make sure Code execution and file creation is on (Settings, then Capabilities; on Team and Enterprise it's under Organization settings). Then go to Customize, Skills, click +, choose Create skill, then Upload a skill, and pick spreadsheet-analyst-skill.zip without unzipping it. On the Free plan and can't find the option? Anthropic's pages disagree on whether Free includes custom skills.

Claude Code

Unzip spreadsheet-analyst-skill.zip into ~/.claude/skills/ (all projects) or .claude/skills/ inside one project, check you have .../skills/spreadsheet-analyst/SKILL.md with no extra folder level, and restart Claude Code.

Using the portable prompt in ChatGPT or Gemini

The file portable-prompt.md carries the same method as copy-paste text. It can't carry the scripts, so it tells the AI to use its own code tool for every number.

Put the Full version in a ChatGPT Project, a custom GPT, or a Gemini Gem so it stays between chats. It's under 8,000 characters, the limit users commonly report for ChatGPT's instruction boxes (OpenAI's own pages don't confirm it). For a one-off, paste the Quick version into a chat, fill in the brackets, and attach your file:

• Upload the file instead of pasting. Pasting loses sheets, date formats, and anything below the visible rows.
• Watch for code. If the AI shows the code it ran, the math is being done properly. If it just answers, treat every number as a draft and check the totals against your sheet.
• In Google Sheets, the Ask Gemini panel turns analysis requests into Python and runs it. It needs an eligible Google Workspace or Google AI plan and works best on native Sheets files (File, then Save as Google Sheets).

Your next step

Give it ten minutes on a file you already know, so you can judge it against numbers you trust:

1. Install it (Section 09) and type "show me a demo". Read the whole Quick Read, verdict included.
2. Drop in one of your own files, ideally one whose totals you've checked by hand, and ask "what does this spreadsheet say?"
3. Compare its total with your sheet's TOTAL row. If they differ, check that it names the rows causing the gap.
4. Read "What this data can't tell you". If it names a column you have elsewhere, add it and ask again.
5. Ask for a report or workbook, open it, and click one Summary figure to see its formula.