The AI Excel Assistant skill for Claude

You download it at the top of this page. You put the zip in your Claude environment, and from that moment on Claude works to this structure. Nothing more needs to happen. What a skill actually is and how that file format works is explained in what are Claude skills NL.

The problem this skill solves is one anyone recognises who has ever pasted in a formula from a Dutch colleague's spreadsheet or a Dutch tutorial. You get back =SOMMEN.ALS(D:D;B:B;"Jansen"), you paste it into your English Excel and a message appears saying the formula contains an error, or that too few arguments have been entered. Two things are wrong at once: the function name is in Dutch while your Excel expects SUMIFS, and the arguments are separated with semicolons while your regional setting expects a comma. Anyone who does not know that goes looking for the mistake in their data while the mistake is in the separator.

Those two things are the core of this skill, and it is exactly what the Dutch-language spreadsheet skills do not do. A skill written for a Dutch audience quietly assumes your Excel is Dutch and that you use semicolons. For most English-speaking office workers that is not the case. That is why this skill delivers every formula twice, and why the separator is the first point in its self-check.

Useful for anyone who is not an Excel analyst but does sit in a sheet every week. The bookkeeper at a small company who wants to see revenue per client. The project lead who has to line up hours from two exports side by side. The online shop owner who has to link a price list to a product list. The sole trader who wants to check their quarterly totals before the tax return goes out. It belongs to the skill library NL we make freely available from our AI and automation service, with no account and no sales email afterwards. That stays the same.

One thing is stated in the file straight away, and we take that over here word for word: the skill cannot open your file and cannot recalculate it. It builds a formula based on what you tell it about your sheet, it explains why it works, and it gives you a check you can run yourself in half a minute. That check is not a politeness but a fixed block in the output, because a formula that looks fine and silently adds up the wrong thing is more dangerous than a formula that throws an error.

The separator. This is the one we see most often, and it is a misunderstanding that keeps doing the rounds. The character between the arguments of a function is not a property of the Excel language but of your computer's regional setting. Excel follows the list separator of the operating system. If your decimal mark is a comma, as is usual in the Netherlands, the comma cannot also separate arguments, so Excel uses a semicolon. On a US setting it is exactly the other way round. In File, Options, Advanced you will find the option Use system separators, but that controls the decimal mark and the thousands separator, not the character between arguments.

The function name. Excel translates function names along with the language of the installation. SUMIFS is called SOMMEN.ALS on a Dutch install, VLOOKUP is called VERT.ZOEKEN, IFERROR is called ALS.FOUT. Paste the Dutch name into an English Excel and you get the error #NAME?, because Excel literally does not recognise the function. It works the same the other way round. Microsoft provides a free add-in for this, the Function Translator, and the skill points to it whenever it is not itself sure of a name. That last part is a deliberate rule: better a detour than a guess.

The version. Not every Excel can do everything. XLOOKUP is available through Microsoft 365 and in the standalone versions from Excel 2021. In Excel 2016 and Excel 2019 it does not exist and you get #NAME?. The same goes for the dynamic array functions FILTER, UNIQUE and SORT. What works perfectly on the author's screen may therefore simply not exist for you, with nothing wrong with your own data. That is why the skill asks the version before it proposes such a function, and delivers a variant with INDEX and MATCH for an older version.

The references. If a formula is correct on row 2 and produces nonsense on row 3, there is almost always a dollar sign missing. Excel shifts references along when you drag a formula, unless you lock them. A dollar sign before the letter locks the column, one before the number locks the row, and both together lock the whole cell. That is why it explains for every formula which reference is locked and why, and tests that by mentally moving the formula one row down and one column right.

The data itself. Numbers stored as text, dates that look like dates but are text, trailing spaces on a product code from an export, client names spelled just slightly differently. This does not throw an error but gives a silently wrong answer, and that is worse. It asks about this beforehand, and puts the pitfalls specific to your sheet in a separate block.

In diff form, with the formula as it appears on the internet and the formula as it works for you.

Not just at the obvious terms like make an excel formula, sumifs, vlookup, create a pivot table or google sheets formula, but also at phrases people actually use: my formula gives #N/A, what does #VALUE! mean, it says too few arguments have been entered, excel will not accept my comma, I have a formula from a video but it does not work, it works on the first row but not once I drag it down, I spend half a day on this every month, I am scared to change anything in this file.

Say something like that to Claude while the skill is loaded and it picks it up automatically. The same frontmatter also lists the cases where the skill should specifically not start, and beneath that sits the licence: MIT, version 1.0.0.

Next comes the core: twelve fixed actions. To begin with it asks what the sheet actually looks like, with sheet names, exact column headers and column letters, and the first and last row with data. It asks the version, because that decides which functions may be used. It asks which separator the formula bar uses and which language the function names are shown in. It rephrases the question first into a sentence you can check, and only then does a formula follow.

That formula it delivers twice, English and Dutch, never mixed. It explains it argument by argument, in the order they appear. It explains separately which references are locked and what happens when you drag the formula. It names the pitfalls specific to this particular sheet. It judges whether a PivotTable or Power Query would need less upkeep. It gives a check you run yourself. It runs the self-check. And it lists every assumption separately at the bottom.

Before those actions start, the skill runs through a mandatory input checklist of ten points: sheet name and column headers with column letter, the first and last row or the name of the table, the Excel version or Google Sheets, the separator in the formula bar, the language of the function names on screen, exactly what needs to come out, where the answer goes and whether the formula gets dragged, what sits in the awkward cells, whether amounts include or exclude VAT and whether there are credit lines mixed in, and how often this comes back. That last point looks like a detail but is not: one-off work calls for a formula, work that recurs every month calls for Power Query.

If something is missing, the skill asks for at most three things at once and then carries on with the assumption stated in the answer.

The output has a fixed structure of ten blocks. What I understand about your sheet. The formula in English. The same formula in Dutch. Explanation per part. References and dragging. Pitfalls in this sheet. PivotTable or Power Query. Check in thirty seconds. Self-check. Assumptions. That order is not arbitrary: the formula comes first so you can paste it straight away, and the explanation sits below it for the moment you want to know why it works. The Assumptions block always closes it, so it stays visible what the answer rests on.

The file also contains two fully worked examples with the user's input and the complete output underneath, written out in all ten blocks. The first is revenue per client per quarter at a fictional plumbing and heating firm, including the VAT warning and the note that client names spelled just slightly differently count as separate clients. The second is a VLOOKUP that gives #N/A at a fictional online shop on Excel 2019, where the answer ends up being INDEX with MATCH because XLOOKUP does not exist there. Both companies are invented and labelled as an example, and no amounts appear that could be read as real.

Lastly a block of writing rules, a list of what the skill never does, a list of cases where a PivotTable or Power Query is the better choice, and the sources. Learning to set up this kind of instruction file yourself is also possible, and that process is laid out step by step in writing a SKILL.md NL.

Microsoft's function documentation is where the names come from. Microsoft publishes a page per function in English, with the English name, the arguments and the versions in which the function is available. The overview is called Excel functions (alphabetical). There is also a free add-in, the Function Translator, that converts a function name between languages. It points there whenever it is not itself sure of a Dutch name, rather than making one up that sounds right.

The separator comes from the regional setting, not from Excel. Excel uses the list separator of the operating system. Where the decimal is a comma, the list separator is a semicolon. That is the whole explanation for the difference between a Dutch spreadsheet and your own screen, and it is the reason the skill has this as its first checkpoint.

Availability per version is documented, not assumed. Microsoft rolled out XLOOKUP from 2019 through Microsoft 365, and in the standalone versions the function is available from Excel 2021. The dynamic array functions FILTER, UNIQUE and SORT followed the same path. Anyone working in Excel 2016 or 2019 does not have them. That is why it asks for the version before such a function ends up in an answer, and for each of those cases a fallback with INDEX and MATCH is already on standby.

Dates are numbers, with one famous quirk. Excel stores a date as a number counted from 1 January 1900, and wrongly treats 1900 as a leap year while doing so. Microsoft describes that as a deliberate choice for compatibility with Lotus 1-2-3. The practical effect on your sheet: you can calculate and compare with real dates, but a date sitting in the cell as text does not take part in a date condition. That is why the skill has you check whether your dates are right-aligned.

The limit of a worksheet is fixed. A sheet has 1,048,576 rows and 16,384 columns, as stated in Excel specifications and limits. Long before you reach that you run into something else: a sheet with tens of thousands of rows plus hundreds of conditional formulas on top of that becomes slow, and then a PivotTable or Power Query is the answer. Power Query has sat in the ribbon by default since Excel 2016, under Data, in Get & Transform Data, and remembers the steps so the next export goes through the same process.

What is notably absent from the source list is a claim about how much faster you will work. You will not find that kind of promise about time or outcome in it. It builds a formula, explains why it works, and puts the check in your own hands. That level-headedness matches how we write about the rest of our work too, for instance in the knowledge base.

After that, using it is simple: describe your sheet in one message. Which sheet, which column letters belong to which header, where the data starts and stops, which version you use, and whether your formula bar wants commas or semicolons. Then say in plain language what you want to know. Have a formula that is broken? Paste the formula itself plus the error message you see, because those two together usually lead to the cause within one step.

Get stuck anywhere while installing, and the full step-by-step plan per environment is at installing Claude skills NL. Prefer to do this with your team rather than on your own, and we take that on together in the AI training NL, and the tools we hang alongside AI besides that are at our MCP server NL and in the overview of tools and plugins.

There is a limit the file itself actively watches for, and it is unusual for a skill: it says out loud when you are better off not using its own answer. As soon as you need two or more breakdowns at once, as soon as the same job comes back every month on a new export, as soon as the file turns slow from thousands of conditional formulas, as soon as several files have to be merged, or as soon as the data has to be cleaned up first, it names the PivotTable or Power Query, even if you explicitly asked for a formula.

That is checkpoint 9 of the self-check and it is not a throwaway line. A formula you have to update by hand every month is not really a solution but an agreement with yourself that you are going to forget.

There is also a category where you are better off leaving it alone. A financial model or annual accounts that an accountant or auditor has to sign off on is not chat work: a human who knows the figures and carries the responsibility should look over it. Tax questions about VAT rates, bases and returns belong with a bookkeeper or tax adviser, not with a formula. A VBA project with several modules is a programming job, not a spreadsheet question. And if the real question is actually about accounting software, a CRM or a database, a formula is the wrong tool, however clever it is.

A limit that often gets forgotten concerns your data itself. As soon as you paste a sheet into a chat window, its contents go along with it. With a list of product codes that is not a problem. With a staff list of names and salaries, a patient list or a complete client base it is different, and then an anonymised example with invented names and amounts is the tidier and safer route. The formula loses nothing by it. It names this itself, too, the moment it sees something like that come past.

Lastly: the skill does not replace understanding. Paste the formula without reading the explanation, and you have a black box sitting in your file that someone else will be staring at next month. That is why the explanation per argument sits in the output by default, not on request. Not sure yet which figures are actually worth measuring before you start adding them up? The Pareto Analyst fits better there, and for the bottleneck in a process the Theory of Constraints Analyst makes more sense than a sum.

The same goes for the monthly round around it: pulling in the export, ordering the figures the same way as last month and lifting out the deviations takes more time than typing the function. That exact round can be prepared for you by an AI employee, with a person checking it before anything goes out the door. Does your team need to get better at this first, together, then that is exactly what the AI training NL is for.