Heidelberg AI Curriculum

T10-L01 · Data & analysis · Level 1 User · 12-minute read, plus 30 minutes of doing

Two thousand rows, one question

You have a table with two thousand rows and one question, and no time to learn pivot tables. A colleague says: "Just give it to the AI." You upload the file, ask your question and get a clean answer with a total, a table and a chart, in twenty seconds.

In a lab it's a plate-reader export: twenty plates, 1,920 wells, one optical density per well. Your supervisor wants the mean signal per sample group by Friday, plus a list of anything odd.

In a company it's the expense export for the year: 2,140 rows from two booking systems. The managing director wants the total per cost centre and the unusually large items before the budget meeting.

The answer arrives quickly, and it looks right. That's the problem. A chat assistant can read part of a file, skip cells it couldn't turn into numbers, or quietly apply a filter it thought you wanted. None of that shows in the tidy answer. This book shows you how to ask so the assistant tells you which rows it used, and how to check that in your own spreadsheet in a few minutes. You are at Level 1 User: the answer is for your own use, so a wrong number costs you a correction, not a decision.

After this you can

Get an answer from a table of a few thousand rows without writing a formula.
Ask for the method and the row count alongside every answer.
Check an answer by recounting one group in your own spreadsheet.
Recognise when the assistant has dropped rows, read numbers as text or applied a filter you didn't ask for.

Before you start

Books: T01-L01 for the four-quadrant method, and T01-L02 for checking output before it leaves your desk.
Claude on the free plan, with code execution switched on. Claude answers table questions by running real calculations on your file, not by reading it like text, and it will show you the steps. ChatGPT's data analysis works the same way and is a fine alternative. If your organisation gives you Copilot in Excel, you can use that instead.
Your spreadsheet program (Excel, Google Sheets or LibreOffice Calc) for the recount.
A table you're allowed to upload. Read T12-L01 first: no personal data about patients, study participants, employees or customers in a personal account. A measurement export or a cost-centre summary without names is usually fine.
About 45 minutes: 12 to read, about 30 to do.

Last verified 23 September 2026. Claude's code execution and file creation is available on every plan, including Free. Switch it on under Settings → Capabilities → Code execution and file creation. Excel files (XLSX) need it switched on; CSV files can be read either way. Attach a file with + at the lower left of the message box, then Add files or photos, up to 20 files per chat. Files for code execution can be up to 30 MB. In Excel, selecting two or more cells with numbers shows Average, Count and Sum on the status bar at the bottom. Numerical Count, Min and Max can be added there too: in Excel for the web, use the arrow next to the last status bar entry; in desktop Excel, right-click the status bar.

The idea in one page

When you upload a table, a good assistant doesn't "read" it. It writes a small program that loads the file and does the arithmetic. That's why the numbers are usually computed correctly. The mistakes happen before the arithmetic, in three places you can't see from the answer:

Silent failure	What happens	How you catch it
1. Rows dropped at import	A second header line, a blank row or a notes section at the bottom makes the program read fewer rows than the file has, or read the notes as data.	Ask how many rows it read. Compare with your own count.
2. Numbers read as text	A cell says `OVRFLW`, `n/a`, `1.234,50` or `12 EUR`. The column becomes text, and the program either skips those cells or stops treating the column as numbers.	Ask which cells it couldn't read as numbers, and how it handled them.
3. A filter you didn't ask for	It "helpfully" leaves out negatives, blanks, last year's rows or a category that looked like a duplicate.	Ask which rows it excluded, and why.

So every question gets three extra requests: how many rows did you read, which rows did you use and leave out, and how did you get the number? Then you recount one small part yourself. If your part matches, the method is probably right. If it doesn't, you've found out before your supervisor or your managing director did.

Two more habits. Words like "outlier", "unusual" and "average" hide choices. Say the rule: "above 1,000 euros", "more than three times the group median", "mean of the wells marked sample". And a chart is only as good as the rows behind it: a smooth chart drawn from 80 per cent of the data looks exactly like one drawn from all of it.

Worked example: three questions, four quadrants

The method comes from T01-L01. Here it is on a plate-reader export, then on an expense export. The same three questions run through both: a total, the outliers, and the missing values.

Lab: the plate-reader export

Structure it. "I have one CSV with 1,920 wells from 20 plates. I need the mean optical density per sample group, a list of wells that look odd, and a list of wells with no value. A useful result: numbers I could put on a slide and defend."

Understand it. She opens the file in Excel first. The last well is on row 1,921, so there are 1,920 data rows under one header. She uploads the CSV to Claude and asks for the mean per group. The answer is clean. Then she asks: "How many rows did you read, and which did you leave out?" Claude reports 1,920 rows read, but 37 wells excluded from the means, because their OD column says OVRFLW. The plate reader writes that when a well is too bright to measure. Those are the brightest wells in the highest-concentration group, so leaving them out quietly pulls that group's mean down. That's silent failure 2. Back to Structure it: "Saturated wells aren't missing values. They're a finding. Report them separately, per plate, and don't average them."

Choose tools. Going down the ladder: the plate reader's own software gives per-plate summaries, but not across 20 plates. A pivot table would do it, but she has never built one. Claude does the calculation, and Excel does the recount. She doesn't need an app or an automation for a one-off question.

Build it. She starts a fresh chat, attaches the file, and sends:

Attached: plate-reader export, one row per well. Columns: plate, well, group, od450.

Before answering, tell me:
- how many data rows you read, and the first and last plate/well
- which od450 cells are not numbers, and what they contain

Then answer three questions:
1. Mean od450 per group, using only numeric cells. For each group give the
   number of wells used.
2. Outliers: wells more than 3 times their group's median. List plate and well.
3. Wells with no value or a non-numeric value, listed by plate, with what the
   cell contains. Do not drop or replace them.

Finally, account for every row: wells used + wells listed in question 3 must
equal the rows you read. Don't apply any other filter. If you had to make a
choice I didn't specify, tell me what it was.

The answer comes back with 1,920 rows read, 1,883 used and 37 listed as OVRFLW, all on plates 4, 11 and 17. It finds two outliers and no empty cells. The row accounting adds up.

Now the recount. She filters the Excel column group to one group, selects its od450 cells and reads the status bar: Count 96, Sum and Average matching Claude's figures for that group to the third decimal. For the high-concentration group, Count says 96 but Numerical Count says 59: the difference is exactly the OVRFLW wells Claude listed there. Her note for Friday reads: "Means from 1,883 of 1,920 wells. 37 saturated wells (plates 4, 11, 17) excluded and listed; these plates need re-reading at a higher dilution."

What it taught her: the most useful line on her slide came from asking which rows were left out. Next time she'll clean the export first (T10-L02).

Company: the expense export

Structure it. "I have 2,140 expense rows for the year. I need the total per cost centre, the unusually large items, and the rows with no amount. A useful result: figures the managing director can rely on."

Understand it. He uploads the file and asks for the totals. They look plausible. He asks how many rows were read and which amounts weren't numbers. Answer: 2,140 rows read, but 312 amounts couldn't be read as numbers. They came from the old travel system, which writes amounts like 1.234,50, with a comma for decimals. The assistant left them out of the totals. That's silent failure 2 again, in a different costume. He also notices that the first answer left out 18 negative amounts, which were refunds. Nobody asked it to: that's silent failure 3. Back to Structure it: "Read both number formats, keep refunds in, and define 'unusually large' as a single item above 1,000 euros."

Choose tools. The same pair as in the lab: the assistant calculates, the spreadsheet checks. He adds one line to the request so the assistant tells him how it converted the comma amounts, because a conversion he can't see is a conversion he can't check.

Build it. He uses the same request with his own columns and three changes: "Amounts may use a comma or a point as decimal separator; convert both and show me five converted examples. Keep negative amounts; they are refunds. Question 2: items above 1,000 euros." The answer lists 2,140 rows read, 2,131 with amounts and 9 empty, 5 conversion examples (all correct), 23 items above 1,000 euros, and the refunds included. The row accounting adds up.

His recount: he filters the spreadsheet to one cost centre and compares Count and Sum with the assistant's figure. The sum in Excel is lower. Count 184, Numerical Count 151: his own Excel also reads the 33 amounts written in the other format as text. He adds those 33 up by hand with a calculator, and the two totals now agree. He writes the finding into his note: "The two systems use different decimal formats; totals include both."

What it taught him: the tool wasn't wrong about the arithmetic. It was wrong about what counted as a number, and so was his own spreadsheet. The next trip round the quadrants is fixing the export from the old system at the source.

In both cases the first answer looked finished. The row accounting is what made it trustworthy.

What goes wrong

Accepting a total without a recount. Symptom: a tidy total goes into a report, and weeks later someone finds it was built on 85 per cent of the rows. Fix: recount one group in your spreadsheet every time. It takes two minutes.

Header not on the first row. Symptom: strange column names like Unnamed: 2, or an instrument name treated as a data row. Fix: tell the assistant which row holds the headers, or delete the lines above it in a copy of the file. Ask for the first and last row it read.

Mixed decimal separators. Symptom: totals that are too low, or amounts a thousand times too small. Fix: say which formats occur, ask for converted examples and check them. In your own spreadsheet, a gap between Count and Numerical Count points at the same problem.

A filter you didn't ask for. Symptom: the total changes when you ask the same question in a new chat. Fix: write "Don't apply any other filter. If you made a choice I didn't specify, tell me" into every request.

A chart built from a subset. Symptom: a clean chart, and a row count in the small print that doesn't match your file, or no row count at all. Fix: ask for the number of rows behind each bar or point before you use the chart.

Uploading personal data. Symptom: a table with patient IDs, names or salaries in a personal chat account. Fix: delete those columns in a copy before you upload, or use the assistant your organisation provides (T12-L01).

Asking the same assistant to check itself. Symptom: "Please double-check" and it confirms its own number. Fix: the check has to come from outside: your spreadsheet, a calculator, or a colleague.

Do it yourself (30 minutes)

Use a table from your own work that you're allowed to upload.

Structure it (5 min). Write three questions: a total, an "unusual" rule you define with a number, and the missing values. Note how many data rows your file has.
Understand it (8 min). Upload the file and ask the first question. Then ask how many rows it read, which cells weren't numbers, and which rows it left out. Note which silent failure, if any, you found. Rewrite your questions if you learned something.
Choose tools (2 min). Decide what you'll recount in your spreadsheet: one group or one category for each question.
Build it (15 min). Start a fresh chat with the full request (row audit, three questions, row accounting, no other filters). Then recount one group per question with the status bar. Write down where you matched and where you didn't, and why.

Exit check

Delivered artifact: three answers from your own table, each with:

the assistant's row accounting (rows read, rows used, rows left out and why);
your own recount of one group from your spreadsheet;
one line: match, or the reason for the difference.

It passes when every source row is accounted for, and each recount either matches or explains the difference.

Rule to remember

Ask it which rows it used.

Ask a Spreadsheet a Question and Check Which Rows It Used