A monthly report that somebody had given up on automating. Two years of columns with most of them hidden, a block of formulas keyed off a range that shifts one column to the right every month, and a budget column that has to be swapped out. Rolling it forward by hand takes about five minutes, and it is five minutes of the fiddliest kind of work there is.

Claude Code had been asked to do it and made a mess of it.

The complaint was not about the answer. The numbers came back correct. What came back wrong was everything that makes a spreadsheet readable: the comma separators, the alignment, the column widths. Thousands and tens of thousands no longer sat under each other, which is the whole reason anyone can scan a column and spot the number that is off.

The instinct in that moment is to prompt harder, or to write the process out in painful detail and try again. Neither was the fix. The fix was to change surface.

Why the same task failed in one Claude and worked in another

Claude Code knows nothing about Excel. It opens a workbook the way any script does, through a Python library, reads it into objects, does the work, and writes a new file out. Everything the library models survives that round trip. Everything it does not model quietly does not.

That is not the model being lazy. It is a file format getting decoded and re-encoded by something that only understands part of it. openpyxl and its relatives carry values and formulas well. Number formats, styles, column widths, merged ranges and conditional rules are a lottery, and the parts that get dropped are exactly the parts nobody notices until they open the file.

The add-in inside Excel does none of that. It drives the application. It writes into cells, edits formulas, sorts, filters, edits pivot tables, applies conditional formatting and creates data validation dropdowns, using Excel’s own operations on the file already open in front of you. Nothing is taken apart, so nothing gets lost in the reassembly.

The same monthly roll-forward, handed to the add-in, worked on the first attempt. Right columns hidden, right columns unhidden, both formula ranges shifted, budget column swapped. It took under a minute and we checked each one on screen.

So if you are grinding at a spreadsheet task and the numbers keep coming back correct and ugly, stop rewriting the prompt. You are using the wrong door. Claude’s office add-ins exist for this, and most people who live in Claude Code never open them.

The unsupported list is two items and neither is the problem

Ask the add-in what it cannot reach and it answers plainly: no data tables, and no macros or VBA. That matches Anthropic’s own Excel documentation exactly. It is a short enough list that you read it, decide neither applies to you, and carry on.

Then you ask it to save next month’s copy of the file, and it fails.

Here is the part that costs people an afternoon. The add-in has no file management capability at all. No save as, no copy, no clone, no rename, no move. It works on the workbook in front of it and has no concept of any other file. That limit is absent from the unsupported list because it was never a capability in the first place, so there was nothing to mark as unsupported.

What happens instead is worse than a flat refusal. Asked for a copy, it goes hunting for a tool that can write a file, finds the Microsoft 365 connector, and tries that. The connector replies that the authorisation grant carries no write scope, so every copy, move and upload will be refused. You are now reading a permissions error.

Two things about that error send you down the wrong path. It describes a permissions problem for something that is not a permissions problem, so people go and raise a ticket with IT. And even with permissions perfect, the connector writes to OneDrive and SharePoint, so a file sitting on your own hard drive was never reachable that way to begin with.

Whether your connector can write at all is a separate question with a separate answer, and it deserves checking on its own terms rather than through an Excel error message.

The practical version: make the copy yourself, open it, then point the add-in at it. Two seconds in Finder buys you the whole workflow.

The settings field that fixes the formatting complaint

The add-in carries its own Instructions field, in the sidebar under Settings. Whatever you put there applies to every conversation you have in that app.

Three things about it are easy to miss and all three matter.

• It is separate from your claude.ai settings. Instructions you set on the website do not arrive here.
• It is separate per Office app. What you write in Excel does not apply in Word or PowerPoint.
• It persists. You set it once, not once per conversation.

This is where “keep numbers in the comma format and keep the columns aligned” belongs. It is also where house rules belong: never change a formula without showing me the before and after, always work on a copy of the tab, never delete a row.

Nobody in the room I was in knew the field existed, and every one of them had the complaint it solves. If you want to work through how this lands across a team rather than one desk, my door is open.

Two smaller behaviours worth knowing while you are in there. The add-in warns before it overwrites existing data, which is the one guard rail standing between a broad instruction and a lost column. And long conversations get compacted, so a session that has run all afternoon is no longer carrying everything you told it at eleven that morning. Restate the rules that matter, or better, put them in the Instructions field where compaction cannot reach them.

Talking to it beat typing at it

The whole monthly process got dictated rather than typed. One continuous spoken instruction covering the file naming convention, which columns hide and which unhide, the two formula ranges that shift and the budget column swap. Roughly a paragraph of conditions that would have been tedious to type and were easy to say.

It parsed cleanly. It also quietly fixed a mis-transcription, reading a spoken “pay special” as paste special from context, without asking and without making a fuss about it.

I am not going to give you a speed multiple, because I did not measure one and the numbers people quote for this do not have much behind them. The more useful observation came from the person doing the dictating: they are more precise when they type and sloppier when they speak. That is the real trade. Voice gets a long specification out of your head in one pass and it will be looser than the version you would have typed. On a first pass, where the alternative was never writing the specification at all, loose and finished beats precise and abandoned.

Then came the sentence that did the most work in the whole hour:

You do not have the full picture here, and I do not know what you are missing either, so ask me questions before you do anything.

It came back with eight of them. Whether the add-backs at the bottom of the statement carry forward. Whether the budget figures stay static for the month or recalculate. What lives on the hidden accounts tab. What gets purged each month and what does not. Whether the version number in the file name increments or restarts.

Every one was a question a colleague would have asked and a prompt would never have contained, because you do not write down the things you have stopped noticing you know. They were answered by voice in a single pass, and the work came out right.

The habit generalises well past spreadsheets. Most bad output from these tools is not the model failing at the task. It is the model doing a decent job on a specification with a hole in it, and the quickest way to find the hole is to make it go looking rather than to guess where it is.

At the end it saved the whole process as a reusable skill, into the working folder, without being asked. That is the first time I have watched somebody produce a reusable asset as a by-product of doing their actual job rather than as a training exercise, and it is a much better way round.

What it costs before it pays

Here is the arithmetic, done out loud by the person doing the work, which is why I trust it more than any number I would have produced.

About forty five minutes to teach the process properly. About five minutes a month to do it by hand across four files. So roughly ten months before the teaching pays itself back.

They said that as a supporter of the approach rather than as an argument against it, and it deserves repeating in that spirit. The first repeatable process you set up costs more than it saves, and it keeps costing more for a long time. The second and third are cheaper, because the conventions, the file layout and the house rules are already shared. That is the actual case for doing it, and saying it up front is the difference between a team that keeps going through month two and a team that quietly stops.

One last thing, and it is the part a compliance team will want before anybody else does. Claude for Excel activity does not appear in Enterprise audit logs and is not covered by the Compliance API. It does not inherit an organisation’s custom data retention settings. The chat history lives in local browser storage rather than on Anthropic’s servers, which also means it does not follow you to another machine. If your governance position is that AI activity on company data has to be reviewable after the fact, the add-in sits outside that today, and it is better to know before you recommend it to a hundred people than after.

None of which changes the recommendation for one person on one workbook. Use the add-in. It is the right tool for the thing it does, it is faster than arguing with a Python round trip, and the limit that will actually bite you is the one that was never on the list you were shown.