Claude for Excel and Spreadsheets: What It Actually Does Well

The one thing worth knowing before anything else: upload the file, do not paste rows. Claude reads the workbook and returns a workbook - not instructions for cleaning it yourself. That removes the step which made most spreadsheet help worthless.

The three jobs it is genuinely good at

1. Cleaning an export

The most common job and also the dullest. A system export arrives with:

• Columns nobody needs
• Dates in three formats
• Amounts stored as text - and therefore not summing
• Header rows in the middle
• Stray spaces in names

Describe what should come out - which columns, in what order, what to do with broken rows - and get the clean sheet back. A quarter-hour job becomes a minute.

2. Reconciling two lists

This is the highest-value job, and it is not obvious.

A bank export against an invoice list. A stock list against a physical count. Customers in one system against another. This is exactly the tedious matching work people do by eye and get wrong - and which needs a lookup formula most people do not write correctly first time.

The right request is not "compare" but "return three sheets: matched, only in list A, only in list B". That structure is what makes the result usable.

3. Turning rows into a report

A thousand transaction rows into a monthly summary by category. A timesheet into a client report. The aggregation itself is easy; what takes time is deciding the structure and building it.

How to phrase a request that comes back right

The difference between a usable result and one that needs redoing is almost entirely in the phrasing.

And two additions that improve almost any request:

• Say what to do with exceptions. A row with no date, an amount that cannot be parsed, a duplicate - drop it or flag it on a separate sheet? Without an instruction, the decision gets made without you.
• Ask for a change rather than starting over. "Add a difference column" works and keeps the context; a new conversation loses all of it.

The three traps that ruin results

1. Dates - the silent failure

This is the most dangerous trap, because it does not show.

In Israel a date is written day/month/year. In a US default it is month/day/year. Which means 03/09 could be the 3rd of September or the 9th of March, and on any day from the 1st to the 12th both readings look completely valid.

What happens in practice: a monthly report that files transactions under the wrong month, and nobody notices because the numbers look plausible.

What to do: state explicitly in the request that dates are day/month. Then check - find a row with a day above 12 and confirm it was parsed correctly.

2. Amounts stored as text

An Israeli export often brings amounts with commas, a currency symbol, or a space before the number. They look like numbers and do not sum.

Say it explicitly: amounts should come out as numbers, with no currency symbol inside the cell.

3. Hebrew in a spreadsheet

Hebrew names in business systems are dirty: a double space, a trailing space, invisible directional characters that arrived via copy-paste. Two names that look identical will not be treated as identical in a match.

That is the number one reason a list comparison returns "not found" for records that exist. Ask for explicit normalisation - stripping stray spaces and invisible characters - as part of the cleaning.

And on formatting: Hebrew works, but directionality in a generated workbook is something to check rather than assume.

The check that takes a minute

A formatted workbook with formulas looks right whether or not it is right. Wrong text looks like text; a wrong spreadsheet looks like a working spreadsheet. That difference is what makes this a habit worth having.

Three checks, under a minute:

1. Count the rows. 1,240 went in, 1,240 came out? If not, why? There may be a good reason, but you should know it.
2. Check one row against the source. Pick a row, verify every field.
3. Check one total. If the grand total matches one you already know, the processing is probably sound.

If those three pass, the rest is probably fine. If one fails, do not fix it by hand - say what is wrong and ask for a correction, because the problem most likely repeats throughout the file.

What not to send

An uploaded file leaves your computer. Usually you can get the same result with far less exposure:

• Delete columns the question does not need - phone numbers, addresses, ID numbers.
• Replace names with identifiers if the names are not needed for the analysis.
• Work on a sample: fifty rows is enough to build the structure, then run it on everything.

When a spreadsheet is the wrong answer

Worth being straight about the boundary.

If the same cleaning repeats every month on the same export, that is automation rather than a conversation. Uploading a file by hand twelve times a year is still twelve manual actions.

And if the whole business runs on spreadsheets - producing spreadsheets faster does not solve that. It produces more spreadsheets, each of them a competing source of truth.

On tool choice: the differences between the tools for business tasks are relevant here, but the quality of the request affects the result more than the choice of tool does.