Can you teach Claude, or any LLM, to follow your standards and your guidelines, so that when it connects to a Power BI model, when it wants to build a semantic model or a report, it does it exactly as you do, following all the instructions that you would normally provide?
Yes, it is possible. And in this article I am going to show you exactly how, using the same simple model I always use in these demos, with a full before and after comparison.
Using AI Tools in Power BI Is Easy. Getting Them to Work Your Way Is a Different Problem (Video: 0:46)
Using AI tools and LLMs is easy these days. You can use MCP Servers, you can use skills, you can use CLIs to do the things you want in a Power BI or Microsoft Fabric environment. I have already covered several of these approaches in past videos.
All of that is possible. There are tools such as MCP Servers, and there are knowledge bases such as skills that enhance the way it works, combined with CLIs, command line interfaces, and together they let you get AI to connect to your model and build whatever you want.
The Big Question, AI Does It, But Not the Way I Do It (Video: 1:51)
But there is a big question people ask me constantly. AI does the things, but it does not do them the way I do them. When I create a measure, I build it under a measure table. I have a set of naming standards. I add a description. When I create a fact table, I have “fact” at the start of the name, or an “f underscore” prefix. I have a set of standards and guidelines that I follow, and AI does not automatically know any of that.
The answer is yes, you can teach AI to do that, and it is genuinely simple.
Here is a simple Power BI Desktop model with FactInternetSales and a Date dimension, related to each other through a relationship on DateKey and OrderDateKey. This model has no sales measure yet, which I want to create by connecting Claude to it, using the Power BI Modeling MCP Server.
Without any skill of my own, without teaching AI that this is my standard way of working with data, I asked Claude to connect to the Power BI model I had open. This uses the MCP Server to create that connection, you do not have to mention MCP Server explicitly, Claude understands it based on what you have installed.
“I want to create three sales measures from this sales amount, total sales, sales YTD, and sales same period last year.”
Notice what I did not say. I did not mention creating these measures inside a measure table. I did not mention any naming standard. I did not mention adding descriptions. My prompt was that plain, and nothing more.
Claude created all three measures, correctly, directly inside the FactInternetSales table. The DAX itself was genuinely accurate, Total Sales was a simple sum, Sales YTD used TOTALYTD based on the date field, and Sales Same Period Last Year used the equivalent time intelligence function. I never told it which date field to use, it figured that out on its own from the relationship already in the model.
This is fantastic. But it is not how I would actually build a Power BI model.
If I were working in a Power BI model myself, I would first create a measure table if one does not already exist. I would create the measures under that table. I would organize them into a display folder. And I would add descriptions to every one of them. None of that happened here, except that Claude did create a display folder on its own this time, which is not guaranteed to happen consistently every time.
The Two Ways to Fix This, and Why One of Them Is Clearly Better (Video: 6:52)
There are two ways to solve this gap.
The first option is enhancing the prompt every single time. Tell Claude explicitly to move the measures into a measure table, add descriptions, follow a naming convention, send a longer, more detailed prompt. This works, but every prompt and every action consumes tokens. Depending on your subscription level with Claude, GitHub Copilot, or any of these tools, you have a limited number of tokens available. Repeating the same detailed instructions in every single prompt is not an efficient way to work.
The second option, and the one I recommend, is teaching AI once using a skill file, so you never have to repeat those instructions again.
A skill is a file you build yourself, containing all the instructions you want an AI tool to follow. If you have used Fabric Skills, these are instructions and guidelines that Microsoft created, and when you pass them to an AI tool like Claude, it follows those instructions.
You can, and honestly should, create your own skill file too. When you do that, you are not only teaching the LLM to follow your instructions, you are also optimizing your token count. With the exact same prompt, you get a meaningfully different, better result.
I started with a genuinely simple plain text file. It did not begin as anything sophisticated, I wrote a few plain instructions:
Always use a dedicated measure table
Organize measures into their relevant display folder
Do not use abbreviations, use fully readable names from the user’s point of view
Add a description to every measure
That is it. I did not hand write a polished, formal skill file from scratch. I gave those simple instructions to AI itself and asked it to compile them into a proper skill file. The result was a file called Measure Table Convention, showing exactly how a measure table should be structured, with concrete good name versus bad name examples, and a template description format.
This is the practical lesson here, you do not need to be a prompt engineering expert to write a skill file. Start with plain instructions in a notepad, and let AI help you formalize them.
Demo 2: The Same Prompt, With the Skill File Loaded (Video: 10:51)
To add the skill, go to Customize inside your AI tool settings, and add a skill. The exact location differs slightly between tools, but it is generally called Skill in every one of them. I uploaded my Measure Table Convention file.
Then, in a brand new chat window, I asked Claude to connect to the same Power BI model. This part behaved identically to before, since my skill file said nothing about the connection itself, so the same MCP Server connection was used.
This time, I sent the exact same prompt as before, word for word. Create three measures from the sales amount, total sales, sales YTD, and sales same period last year.
Claude Detects and Applies the Skill Automatically (Video: 11:45)
Before doing anything else, Claude reported that the Measure Table Convention skill was loaded. I never told it to use that skill in this prompt, it detected and applied it automatically.
Because there was no measure table yet, Claude created one first. Then it created all three measures under that table, inside the correct display folder, using full, non-abbreviated names exactly as instructed, and every single measure had a proper description attached, visible on hover in the field list.
Same prompt. Same token count spent on the actual request. Completely different, correctly structured output, because the standards lived in the skill file instead of being repeated in the prompt.
Why This Matters for Anyone Using AI to Build Power BI Solutions (Video: 14:27)
If you say “I am going to use Fabric Skills” or “I am using a community published skill,” that alone is not enough. You need to customize it, or add your own skill on top of it, because the way you work is likely different from the generic guidance in a shared skill file. Your organization’s naming standards, your measure table conventions, your display folder structure, these are specific to you.
This is not tool specific either. It does not matter if you use Claude, GitHub Copilot, Codex, or any other AI assistant, they can all accept a skill file, and using one means fewer prompts sent while achieving better, more consistent results.
You are optimizing your token usage, and you are creating a model that looks much closer to what you would build yourself, if you had done it manually.
Key Takeaways
Without a skill file, AI produces technically correct but generically structured Power BI models, no measure table, no consistent naming, no descriptions by default
A skill file is simply a text file containing your standards, you can start with plain, informal notes and let AI help formalize it into a proper skill document
The same prompt, with a skill file loaded, produces a model that follows your exact conventions automatically, without repeating instructions every time
This works across AI tools, Claude, GitHub Copilot, Codex, or any assistant that accepts a skill file
Using your own custom skill on top of any shared or community skill, like Fabric Skills, is important, because your organization’s standards are specific to you
This approach genuinely optimizes token usage while producing results closer to what you would build by hand
Reza Rad is a Microsoft Regional Director, Data Platform MVP (16 consecutive years), Author, and Trainer. He is the co-founder of RADACAD and the author of multiple books on Power BI, Power Query, and Microsoft Fabric. You can follow him on LinkedIn and subscribe to the RADACAD YouTube channel.
Reza Rad is a Microsoft Regional Director, an Author, Trainer, Speaker and Consultant. He has a BSc in Computer engineering; he has more than 20 years’ experience in data analysis, BI, databases, programming, and development mostly on Microsoft technologies. He is a Microsoft Data Platform MVP for 12 continuous years (from 2011 till now) for his dedication in Microsoft BI. Reza is an active blogger and co-founder of RADACAD. Reza is also co-founder and co-organizer of Difinity conference in New Zealand, Power BI Summit, and Data Insight Summit.
Reza is author of more than 14 books on Microsoft Business Intelligence, most of these books are published under Power BI category. Among these are books such as Power BI DAX Simplified, Pro Power BI Architecture, Power BI from Rookie to Rock Star, Power Query books series, Row-Level Security in Power BI and etc.
He is an International Speaker in Microsoft Ignite, Microsoft Business Applications Summit, Data Insight Summit, PASS Summit, SQL Saturday and SQL user groups. And He is a Microsoft Certified Trainer.
Reza’s passion is to help you find the best data solution, he is Data enthusiast.
His articles on different aspects of technologies, especially on MS BI, can be found on his blog: https://radacad.com/blog.
