Power BI Portfolio Project: Build One With Claude (No Data)

Power BI portfolio project · 8 steps with Claude

Zero to portfolio

A Power BI portfolio project shows what you can actually build, and you do not need real data to make one. In this guide, I build a complete sales report from zero with Claude: sample data with seasonality, a star schema semantic model and a report with three pages. All I start with is an empty Power BI file and two prompts.

I followed every step in my YouTube video, so you can watch the build and read the steps side by side. Moreover, I share the exact prompts and the things that went wrong, because the first version was not perfect. If you want to build a Power BI report with AI for the first time, this is a good project to start with.

The full build on YouTube: from an empty file to a finished report.

Key takeaways

You do not need company data for a Power BI portfolio project. Claude can write realistic sample data with seasonality for you.
The Power BI Modeling MCP server builds the semantic model, and the powerbi-report-cli skill tells Claude how to build the report.
Two Power BI Desktop settings matter: Desktop Bridge and “Detect and reload external PBIP changes”.
Ask for an HTML mockup first, so you can fix the design before anything is built.
The first version is a draft. In my case, one round of concrete feedback already fixed most of it, but a real portfolio report needs more iterations than this short concept.
The same steps work for any data analyst portfolio topic, from sales to finance or marketing.

What is a Power BI portfolio project?

A Power BI portfolio project is a finished report you build to prove a skill. It has a data model, measures and pages that answer real business questions. In other words, it is something a hiring manager or client can open and click through, instead of a line on your resume.

A portfolio shows what you can build. A resume only says that you can.

It worked for me, too. When I applied for a job many years ago, I built a small portfolio project: a cohort analysis. I showed it to the client and got hired. As a result, I always recommend building one report that fits the problem of the company you want to work with.

However, most people get stuck on the data. You cannot show the reports you built at work, because they contain client and sensitive information. Public datasets are an option, but everyone uses the same ones. That is the problem this guide solves.

What you need: the Claude Code Power BI setup

You need three things: Power BI Desktop, a coding agent like Claude Code, and two Microsoft tools that teach the agent Power BI. Both Microsoft tools are free on GitHub. Together, they are my whole Claude Code Power BI setup for this build.

Tool	What it does	Do you need it?
Power BI Desktop	Holds the file, loads the data and shows the report.	Yes
Claude Code	The coding agent that writes the data, the model and the report files.	Yes. Any coding agent works, but I use Claude Code.
Power BI Modeling MCP server	Connects Claude to your semantic model: tables, relationships, measures, refresh and DAX queries. It cannot build report pages.	Yes, for the semantic model.
semantic-model-authoring skill	Written instructions for building a good semantic model.	Optional. The MCP server can build the model on its own.
powerbi-report-cli skill	Written instructions for planning, designing and building report pages and visuals in a PBIP.	Yes, for the report.

Power BI Modeling MCP server or skill: what is the difference?

In short, the MCP server is the hands and the skill is the recipe. The Power BI Modeling MCP server is a Microsoft tool that connects an AI agent to a Power BI semantic model in Desktop, in a PBIP folder or in a Fabric workspace. According to its README, it can only run modeling operations and cannot change report pages. A skill is a set of written instructions. It cannot write files on its own; instead, Claude reads it and writes the files following it.

Microsoft lists Claude Code, GitHub Copilot CLI, Cursor and Codex among the agents that skills-for-fabric supports. Specifically, you can find the semantic-model-authoring and powerbi-report-cli skills in Microsoft skills-for-fabric, and the server in microsoft/powerbi-modeling-mcp. For this guide, I assume the MCP server is already installed. I also wrote a separate post on what Power BI Report CLI does.

Microsoft skills-for-fabric repository on GitHub with the skills folder Claude uses to build reports — The skills folder in Microsoft skills-for-fabric. powerbi-report-cli is the one you need for the report.

Step 1: Start your Power BI portfolio project with an empty PBIP file

Open Power BI Desktop, create a blank report and save it as a Power BI project (PBIP). A Power BI project (PBIP) is a way to save a report as plain text files, with one folder for the report and one for the semantic model. As a result, Claude can read and write every part of the file.

Claude can also create the file for you. Still, I find it easier to start from a saved blank file, because it is less prone to errors. I saved mine as “Sales Portfolio Example”. At this point it has no data, no tables and no pages. That is the starting point.

Two limits from the Microsoft PBIP documentation are worth knowing. First, you cannot save a PBIP directly to OneDrive or SharePoint. Second, project paths are limited to 260 characters on Windows by default, so use a short local folder.

If PBIP is new to you, read my comparison of PBIX vs PBIR vs PBIP.

Saving an empty Power BI Desktop file as a Power BI project PBIP — Step 1: save the blank file as a Power BI project (PBIP).

Step 2: Turn on two Power BI Desktop settings

Two settings let Claude work with your open file. Both live under File, Options and settings, Options.

Desktop Bridge (Global, Security): “Enable external tool access to Power BI Desktop through secure local APIs”. This lets Claude talk to Power BI Desktop and take screenshots of your report, so it can check its own work for errors and misaligned visuals.
Detect and reload external PBIP changes (Global, Preview features). Claude only changes the JSON files in the background. With this setting on, Desktop asks you to apply those changes, so you do not have to close and reopen the file.

I explain the first one in detail in my post on the Power BI Desktop Bridge.

Desktop Bridge option enabled under Global, Security in the Desktop options — Desktop Bridge: lets Claude reach Power BI Desktop and take screenshots.

Preview feature Detect and reload external PBIP changes turned on — Preview features: detect and reload external PBIP changes.

Get the free guide: how I build Power BI reports with AI

Everything from this build written down: the skills, the prompts and the whole setup, plus a PBIP file to try it yourself.

Step 3: Write one prompt for the data, the model and the mockup

One prompt does most of the work. I split it into two parts: first the data and the semantic model, then a mockup of the report. Claude must stop and wait for my approval before it builds anything in the report.

I used a fictional outdoor-gear company, because sales is the easiest topic to understand. Here is the exact prompt from the video.

Prompt 1, part 1: the data and the semantic model

Show prompt 1, part 1

I want to build a sales dashboard for a fictional outdoor-gear company. We start with the data and the semantic model, and finish with a mockup of the report.

Part 1: data and semantic model
1. Suggest the typical KPIs for a sales dashboard first.
2. Write CSV files for 2 years of daily orders: sales, products (6 categories), countries (15 countries in 3 regions) and a calendar. Add real seasonality: a slow January and February, a summer outdoor season, Black Friday and a December peak. Flip the seasons for Australia and Brazil. Launch a new product category in March of year 2.
3. Build the semantic model in my open Power BI file with the Power BI Modeling MCP, following the semantic-model-authoring skill: load the CSVs with Power Query, add a date table, relationships, and one measure per KPI in a '!Measures' table.
4. Follow best practices: a star schema with one fact table, single-direction one-to-many relationships, hidden key columns, a marked date table, sort-by columns for month and weekday, and explicit measures only.

Prompt 1, part 2: the report mockup

Show prompt 1, part 2

Part 2: the report mockup (only when the model is done)
5. Create an HTML mockup of the report, with the same look and colors as the attached example.
6. Canvas size 1600 x 900 pixels, with a sidebar on the left for navigation.
7. 3 pages, from high level to detail:
   - Overview: high-level KPI cards and the main trend
   - Drill-down: revenue by region, then country, then product category
   - Analysis: seasonality, growth versus last year, online versus retail
   And a hidden drill-through page for one country or product.
8. Dynamic titles that show what is selected (for example "Revenue for Germany").
9. Every visual on a background with rounded corners of 10 pixels, and the same spacing between all visuals.
10. Use only the KPIs and measures from the semantic model.

Show me the mockup and wait for my approval before you build the report.

Why the best practices are in the prompt

Notice point 4. I spell out the best practices I want, such as one fact table, single-direction relationships and explicit measures. As a result, the model follows them from the start, and I do not have to fix them later.

The Power BI portfolio project prompt in Claude Code — The full prompt in Claude Code, with an example design attached.

Step 4: Review the KPIs and the HTML mockup

Claude first suggests the KPIs, then shows an HTML mockup of the report. An HTML mockup is a web page that looks like the finished report. It opens in any browser, it is easy to share, and its filters even work, unlike a static Figma file.

For the look, I attached a design I liked from Dribbble (shout out to Nasim): rounded corners, a sidebar for navigation and the same colors. For example, you can also give Claude a dashboard you built before and say “make it like this”. My guide on dashboard design tools and inspiration has more sources.

My first mockup was good, but not right. The drill-down page was not a real drill-down, and the analysis page was third. So I asked for a hidden drill-through page with order details, including the order number, country and customer. Claude noticed that the data had no customer table yet and asked to add one. After that round, the report had two pages plus one hidden drill-through page.

HTML mockup of a Power BI sales report created by Claude before the build — The HTML mockup: I approved the design here, before a single visual existed.

Step 5: Get Power BI sample data with real seasonality

Good Power BI sample data needs patterns, not just rows. For this build, Claude wrote CSV files for sales, products, countries, customers and a calendar. In my run, the sales table had about 248,000 rows and the product table 46 products. That is enough to make a report feel real.

Seasonality is the part that makes the data believable. Flat random numbers smell like fake data. Instead, I asked for a slow January and February, a summer season, Black Friday and a December peak, with flipped seasons for Australia and Brazil. Consequently, the trend charts show patterns you can actually talk about in an interview.

Data source	Good for a portfolio?	Why
Your company reports	No	Client and sensitive data. You cannot show it.
Public datasets (for example Kaggle)	Partly	Free and real, but many people use the same ones, and you cannot shape the story.
Sample data written by Claude	Yes	You choose the topic, the size and the patterns, such as seasonality and a new product launch.

I tried a ChatGPT version of this back in 2024, in my post on creating a Power BI dataset with ChatGPT. The difference now: Claude writes the Power BI sample data, loads it and builds the model in one go.

CSV files with sample sales data written by Claude: sales, calendar, countries and products — The data folder Claude created: one CSV file per table.

Step 6: Build the semantic model with the Power BI Modeling MCP

A semantic model is the layer between your data and your report: the tables, relationships and DAX measures that every visual reads from. In this step, Claude used the Power BI Modeling MCP server to load the CSV files with Power Query and build the model in my open file. The semantic-model-authoring skill was an extra reference. You can probably use only one of them, but together they work like a superpower.

A star schema is a model layout with one fact table of events, such as orders, in the middle and dimension tables, such as products or dates, around it. As a result of the prompt, I got a clean one. The Sales fact table sits in the middle, with four keys to Customer, Country, Product and Date. The key columns are hidden, and all measures live in a separate ‘!Measures’ table. Microsoft explains why this layout matters in its star schema guidance.

It did not work the first time, though. When I opened Power BI, I saw no tables, no measures and no relationships. So I told Claude exactly that, and it added them through the MCP server. Then I clicked Refresh once, because Power BI has to load the CSV files the first time.

Star schema semantic model in Power BI built with the Power BI Modeling MCP server and Claude Code — The star schema: Sales in the middle, four dimensions and a measures table.

Step 7: Build a Power BI report with AI using powerbi-report-cli

Once the model was right, I sent a second, shorter prompt. It tells Claude to build the approved mockup in the same project with the powerbi-report-cli skill:

Show prompt 2

The mockup looks good. Now build the report in the same Power BI project with the powerbi-report-cli skill, exactly like the approved mockup.

1. Same canvas size, sidebar navigation (only the visible pages), pages, spacing and rounded corners.
2. Hide the drill-through page, and set it up so you reach it by right-click > Drill through from the drill-down and analysis pages.
3. Make the dynamic titles with measures in the '!Measures' table.
4. Use only measures from the '!Measures' table in the visuals.
5. When it's done, validate the report and show me a screenshot of every page.

Claude then wrote the theme, the pages and the visuals as JSON files, validated the report, reloaded Desktop and took screenshots to check its own work. Meanwhile, Desktop showed an “apply external changes” message. That is the reload setting from Step 2 doing its job.

The full build took almost 28 minutes. However, part of that was my fault: I missed an approval prompt and Claude waited. A single page or a small change is much faster.

First version of the Power BI report built by Claude with the powerbi-report-cli skill — The first version: close to the mockup, but the black was too strong and one chart was missing.

Step 8: Review and iterate on your Power BI portfolio project

The first version looked good at first glance. The page navigation worked, the dynamic titles changed with the selection, and right-clicking United States opened the order details page with a back button. Still, there were problems. The black was too strong, the “year at a glance” visual did not work, a trend chart from the mockup was missing, and the visuals highlighted instead of cross-filtering.

So I gave Claude that exact feedback. After one round, the black was gone, the colors were a brighter yellow-green and the year at a glance worked. Filtering by a month changed the numbers, and the dynamic title changed to the region I selected.

The first version is a draft, not the result. Plan for a few rounds of feedback, and make every round specific.

Here are the three final pages. The order details page is hidden and only opens through drill through. Microsoft explains the setup in its drill through documentation, and my guide on drill through vs drill down shows when to use which.

The build in numbers

What	In my build
Prompts to build it	2, plus follow-up messages for the mockup, the missing tables and the report feedback
Sample data	2 years of daily orders, about 248,000 sales rows, 46 products in 6 categories, 15 countries in 3 regions
Semantic model	1 fact table, 4 dimension tables (Customer, Country, Product, Date) and a measures table
Report	2 visible pages and 1 hidden drill-through page, on a 1,600 x 900 pixel canvas
Report build time	Almost 28 minutes, including a few minutes waiting for my approval

Final overview page of the Power BI portfolio project with KPI cards, revenue trend and top countries — Overview page: KPI cards, revenue vs last year and the top countries.

Analysis page with seasonality by hemisphere and growth versus last year — Analysis page: seasonality by hemisphere, growth vs last year, online vs retail.

Hidden drill-through page with order details for one country — Hidden drill-through page: every order line for one country, with a back button.

This is not where I would stop

To be clear, these pages are not a finished report. They show the result after only a couple of iterations, and I would not stop here for a real portfolio. This build is a short concept: it shows you how the process works, not how polished the end result can get.

In reality, you would run more rounds of feedback. Over time, you would also build a library of what you like: how you want to visualize certain things, which chart types you prefer, and what the background should look like. Then Claude starts closer to your style every time, and each new Power BI portfolio project needs fewer rounds.

7 lessons learned from my Power BI portfolio project

These lessons come from what actually happened in the video, not from theory.

Lessons 1 to 4: review the work

Check every measure. Do not trust what Claude builds blindly. For example, I found a “Gross Revenue” measure built as sales plus discount, so I had to check how the discount is stored. Likewise, the analysis page showed a “trend index” I had never asked for. Open each measure and read how it is defined.
The first version is a draft. It is rarely perfect on the first try. Claude has a good idea, but you refine it, just like on a real project.
Give concrete feedback. “Make it look better” does not work. “The black is too strong, use cross-filter instead of highlight, the year at a glance does not work” does. Likewise, if you want it to look more like Power BI, show Claude a real Power BI example.
Mockup first. The HTML mockup took minutes to change. In contrast, the same change in a finished report would have taken a full build.

Lessons 5 to 7: work with Claude and Desktop

Use Claude Code over Cowork for screenshots. In my tests, only Claude Code worked with the Desktop Bridge, so only Claude Code could take screenshots and check its own work. It did not work for me in Cowork sessions.
Restart Desktop when it gets stuck. After one build, Power BI Desktop got stuck in a strange state. I asked Claude to close and reopen it, and it restarted Desktop on its own. Then I closed the recovered files window and carried on.
Build a style guide for next time. Once the feedback rounds are done, write down how you like things: colors, spacing, titles, interactions. In the next session, the skill reads it, so your next report starts closer to the result.

One bonus trick: you can ask Claude to change the PNG background image of a page to match your KPIs and visuals, so you do not need Figma for that. I cover backgrounds in my post on the Power BI background SVG.

Claude writes the data and the model in minutes. You still have to check every measure, because you are the one who presents it.

How to turn one project into a data analyst portfolio

This approach is not only for sales. It works just as well for finance, marketing or HR. The trick is to pick the problem first: if you apply for a job, build a report that fits that company and shows how it could help them.

That is how you show value from the start, and it sets you apart from people who only send a resume. Furthermore, a good portfolio project answers clear business questions. My post on Power BI report structure shows the four questions every report should answer, and my Power BI interview questions help you explain your choices afterwards.

What makes a strong Power BI portfolio

A strong Power BI portfolio is small and focused. One report that solves a clear problem says more than five copies of the same tutorial. For each project, show the report, the semantic model and a short note on the business question it answers. In addition, add the HTML mockup, because it shows how you plan.

Why this is a good Power BI project for beginners

Sales is the easiest topic to start with, because everyone understands revenue, orders and margin. Moreover, you do not have to write the DAX or Power Query yourself. However, you do need to read them, so you can check what Claude built. That makes it a good Power BI project for beginners who learn by reviewing.

One last word of caution: review every measure before you show your data analyst portfolio to anyone.

Want the whole setup on paper?

My free PDF guide explains the skills, the prompts and how I get from a basic report to one with dynamic elements and a good user experience.

Get the free guide →

FAQ

Questions about Power BI sample data

How do I build a Power BI portfolio project if I have no data?

Let an AI agent write the data for you. In this guide, Claude wrote CSV files for two years of daily orders, with products, countries, customers and a calendar. In addition, it added real seasonality, so the data does not look fake.

Is it okay to use AI-generated sample data in a data analyst portfolio?

Yes, as long as you say so. A portfolio shows how you model data, write measures and design a report. Fictional data is fine for that, and it avoids sharing confidential company data.

Questions about the Claude Code Power BI tools

Do I need both the Power BI Modeling MCP server and the skills?

Not strictly. The MCP server can build the semantic model on its own, and the semantic-model-authoring skill is only an extra reference. However, the MCP server cannot build report pages, so for the report you need the powerbi-report-cli skill.

Can I use Claude Cowork instead of Claude Code?

You can try, but in my tests only Claude Code worked with the Power BI Desktop Bridge. As a result, only Claude Code could take screenshots of the report and check its own work.

How long does it take Claude to build the report?

In my run, the full report took almost 28 minutes, including a few minutes where Claude waited for my approval. A single page or one small change is much faster.

How many projects do I need in a Power BI portfolio?

One strong project is enough to start. After that, add one project for each area you want to work in, such as finance or marketing. In short, quality beats quantity: every project should answer a clear business question.

What is a good first Power BI project for beginners?

A sales report is the easiest start, because everyone understands revenue, orders and margin. After that, build one that fits the company or industry you want to work in, such as finance, marketing or HR.

Last updated: 26 October 2026. Written by Lukas Reese, Power BI consultant, from his own build in the video above.