How To Use Claude Code for Power BI Development

If you use AI in your Power BI work today, the loop probably looks like this: you ask ChatGPT or Claude for a DAX measure, copy the answer into Power BI Desktop, check that it works, and go back for a fix or the next measure. The AI writes the code, and you carry it between two windows, every time.

An agent does the carrying. Claude Code from Anthropic and Codex from OpenAI run in a terminal, read your files, make the change you describe, and ask before they act. This guide walks through my setup as it was when I recorded the video in August 2026: Claude Code in the terminal inside Visual Studio Code, Microsoft’s Skills for Fabric, and a Power BI project. One plain-English prompt adds a New Customer Revenue measure to the model. The rest of the guide covers what keeps you in control of a tool like that: permission prompts, permission modes and skills.

The same measure two ways: in the chat loop you move the code, and the agent edits the file after asking you.

This works on Power BI because of how a Power BI project is saved. Save a report as a Power BI project (PBIP), and the semantic model and the report become folders of text files, with the model written in TMDL. An agent that can edit text can edit those files. Power BI as Code covers the format.

0:21What a terminal is, and where to find one

A terminal is a text box where you type an instruction and your computer does something. Windows has one called PowerShell, and a Mac has one called Terminal.

The terminal I like best is the one inside Visual Studio Code. VS Code is a text editor with a strong terminal built in. Open it with Ctrl+J on Windows or Cmd+J on a Mac; VS Code’s keyboard shortcut sheet also lists Ctrl+` for showing the integrated terminal. When you open a folder with File > Open Folder, the terminal starts in that folder.

Two commands to start with:

ls
clear

ls lists what’s in the current folder. In the demo that folder is C:\Demo\Sales, a Power BI project, and ls shows its folders and files. It goes one level deep: it lists the folders but not what’s inside them. In PowerShell, ls is another name for Get-ChildItem, and Get-ChildItem -Recurse lists the subfolders’ contents too. clear clears the screen.

ls lists one level of the project folder: the folders and files directly inside it.

You could ask a chat bot how to write a command and paste its answer into the terminal. Or you can let the AI write the commands and run them itself.

1:39Claude Code and Codex: agents that run the commands

The program that lets an AI model act on your machine, running commands and reading and editing files, is called a harness. Anthropic’s is Claude Code, and OpenAI’s is Codex. Claude Code most naturally uses Anthropic’s Claude models, and Codex most naturally uses OpenAI’s.

Company	Chat window	Agent (harness)
Anthropic	Claude	Claude Code
OpenAI	ChatGPT	Codex

This guide follows my setup: Claude Code, running Claude. Chat answers. Agents act. recaps the difference between the two columns on one page.

2:26Starting Claude Code and the first prompt

You install Claude Code once, from a terminal. Anthropic’s setup page gives these commands:

# Windows PowerShell
irm https://claude.ai/install.ps1 | iex

# macOS, Linux or WSL
curl -fsSL https://claude.ai/install.sh | bash

You can also install it with winget install Anthropic.ClaudeCode on Windows or brew install --cask claude-code on a Mac. When the installer finishes, open a new terminal and run claude --version. A working install prints a version number; the demo ran 2.1.247.

Claude Code needs a Claude Pro, Max, Team or Enterprise plan, or a Console account. The free Claude plan doesn’t include it. On Windows, Git for Windows is optional: without it, Claude Code runs its commands in PowerShell.

To start a session, open the project folder in VS Code with File > Open Folder, open the terminal, and type claude. The first time, it signs you in through your browser. When I typed claude, the terminal suggested adding --permission-mode bypassPermissions. I left it off, and the permission modes section below explains why.

Add Microsoft’s Skills for Fabric

Microsoft’s Skills for Fabric are a critical part of my setup. Microsoft released them at Build in June 2026 as an open-source collection of skills for AI agents, installed as plug-ins from Microsoft’s marketplace. With Claude Code running, type these at its prompt:

/plugin marketplace add microsoft/skills-for-fabric
/plugin install fabric-skills@fabric-collection
/plugin install powerbi-authoring@fabric-collection

The first command adds Microsoft’s marketplace. The other two install the plug-ins: fabric-skills, the main bundle for Fabric, and powerbi-authoring, the Power BI one. Then quit Claude Code and start a new session, so it loads the skills. The demo ran version 0.3.4 of powerbi-authoring.

Besides skills, powerbi-authoring includes the Power BI Modeling MCP server, a program that gives the agent actions it can run against a semantic model. The measure later in this guide was written through it. How to Install Microsoft’s Fabric Skills tours the whole repository, and How To Use Microsoft’s Fabric Skills To Build a Power BI Report builds a full report with the Power BI plug-in.

The first prompt

The first prompt in the demo was a question about the folder:

Please tell me about the contents of this folder.

It’s the question ls answers, and the answer went much further. Claude Code wrote a PowerShell command to list every file, asked my permission to run it (the next section reads that command), and in all “searched for 1 pattern, read 7 files, ran 3 shell commands”. Then it described the project: “a demo “Sales Performance” report built from CSV data”. Its summary included:

• Data/: seven source CSV files forming a star schema, with Sales.csv as the fact table (2,000 rows, January 2024 to December 2025) and six dimensions: Date, Customer, Product, Store, Employee and Promotion.
• Sales.SemanticModel/: the data model in TMDL, with relationships from Sales to all six dimensions and 16 measures in tables/Sales.tmdl.
• Sales.Report/: the report definition in PBIR format, with four pages.

3:39Permission prompts: yes, always, or no

Before Claude Code runs a command, it shows you what it’s about to do and waits. This is the command it wrote to list the folder:

Get-ChildItem -Recurse -Force | Where-Object { -not $_.PSIsContainer } | Select-Object @{n='Path';e={$_.FullName.Replace('C:\Demo\Sales\','')}}, Length | Format-Table -AutoSize

Read it left to right, one step at each pipe (|):

1. Get-ChildItem -Recurse -Force lists everything in the folder and all its subfolders, hidden items included.
2. Where-Object { -not $_.PSIsContainer } keeps the files and drops the folders.
3. Select-Object keeps two columns: each file’s path with C:\Demo\Sales\ taken off the front, and its size (Length).
4. Format-Table -AutoSize prints the result as a table.

It’s a specific script that I wouldn’t want to write myself, and you don’t have to. You need to read it well enough to see that it only lists files. Claude Code also flagged it: “Command contains script block that may execute arbitrary code”. A script block in PowerShell is code inside braces, and this command has two.

The permission prompt for the folder listing: the command, a warning, and three answers.

You have three answers:

• Yes runs it this once.
• Yes, and don’t ask again runs it and remembers the approval. Anthropic’s permissions page says that for a shell command the approval is permanent, for that repository and that exact command. For a file edit, it lasts until the session ends.
• No doesn’t run it.

The line underneath offers three more keys: Esc to cancel, Tab to amend, and Ctrl+E to explain. I rarely choose the second answer, because the agent seldom asks to run exactly the same command twice.

Not everything prompts. Claude Code reads files in the folder you started it in without asking, and a built-in set of read-only commands runs without a prompt too. File edits and other commands ask.

Once you answer yes, the command runs, its result goes back to the model on Anthropic’s servers, and the model reads it. It may run more commands before it writes up what it found.

4:53One prompt adds the DAX measure to the model

Then the real job, in one prompt:

Add a measure called New Customer Revenue: revenue from customers whose first-ever purchase happened in the selected period. Put it in the Sales table and format it as currency.

It names the measure, states the business rule, says which table the measure goes in, and sets the format.

Claude Code worked for about a minute and 50 seconds (“Cogitated for 1m 50s”), asking permission as it went. Then it reported that “the only change across all 16 TMDL files is the new measure block in Sales.tmdl”. It named the measure, Sales[New Customer Revenue], and its format, $#,##0 (“matches Revenue and the other currency measures”), and saved it at line 88 of Sales.tmdl, the Sales table’s file in Sales.SemanticModel\definition\tables, with nothing copied or pasted. This is the DAX it wrote:

VAR _PeriodStart = MIN ( 'Date'[Date] )
VAR _PeriodEnd = MAX ( 'Date'[Date] )
VAR _NewCustomers =
    FILTER (
        VALUES ( 'Sales'[Customer ID] ),
        VAR _FirstPurchase =
            CALCULATE (
                MIN ( 'Sales'[Sale Date] ),
                ALLEXCEPT ( 'Sales', 'Sales'[Customer ID] )
            )
        RETURN
            _FirstPurchase >= _PeriodStart
                && _FirstPurchase < _PeriodEnd + 1
    )
RETURN
    CALCULATE ( [Revenue], KEEPFILTERS ( _NewCustomers ) )

The measure it matched the format to is the first one in the same file:

	/// Net sales revenue after discounts. Primary measure of the report.
	measure Revenue = SUM(Sales[Net Amount])
		formatString: $#,##0
		lineageTag: c1c52f36-46b8-4d95-944b-14c50e42cb3a

Review it like a colleague’s work

The agent wrote the measure, and deciding whether it’s right is still your job. Its summary listed the design decisions behind it, which makes the review easier:

The agent’s design decisions and its first caveat, from the end of its summary.

• The model has a Customer[Join Date] column, and the agent didn’t use it: “a join date isn’t necessarily a purchase date”. “First-ever” comes from the Sales fact instead, as each customer’s earliest Sale Date.
• ALLEXCEPT ( 'Sales', 'Sales'[Customer ID] ) strips every filter except the customer when it looks for that first purchase, so a product or region slicer can’t turn an existing customer into a new one.
• “Selected period” is the earliest and latest 'Date'[Date] in the current filter context, so the measure works by year, quarter, month or a custom range. With no date filter, every customer counts as new for the full calendar.
• < _PeriodEnd + 1 rather than <= _PeriodEnd keeps the end-of-period check right if Sale Date ever carries a time.
• Revenue for the new customers is then filtered to the selected period through the relationship from Sales to Date.

Microsoft Learn’s pages on ALLEXCEPT and KEEPFILTERS explain the two functions in full.

The caveat matters most. The agent “could not execute the measure against data”, because no Power BI Desktop instance was running. The DAX refers only to columns and measures it checked exist, but at that point the measure had never returned a number. Its suggested next step was to open Sales.pbip in Desktop and check the values, and that’s the step to take before you trust the measure.

The caveat also says how the agent edited the file: “The model was authored via the MCP against the PBIP folder offline”. It used the Power BI Modeling MCP server from the powerbi-authoring plug-in and worked on the project’s files with Desktop closed. With the project in Git, the change is also a diff you can read and revert; Get Your Power BI Reports AI-Ready with GitHub sets that up.

5:34Permission modes: manual, accept edits, plan, auto, bypassPermissions

Claude Code asked before almost every step of that job. That’s safe, but it doesn’t save much time if you have to sit and watch it. The permission mode sets what it may do without asking. The status bar under the prompt shows the current mode, and Shift+Tab cycles through them. The demo’s session was in Manual mode (“manual mode on”), and each press of Shift+Tab moved it to accept edits, then plan, then auto, then back to Manual.

Mode	Status bar	What runs without asking
Manual (default)	manual mode on	Reads only
Accept edits (acceptEdits)	accept edits on	Reads, file edits, and common filesystem commands such as mkdir, mv and cp
Plan (plan)	plan mode on	Reads. Claude explores and writes a plan, and doesn’t edit until you approve the plan
Auto (auto)	auto mode on	Everything, with a second model, a classifier, reviewing actions in the background
Bypass permissions (bypassPermissions)	bypass permissions on	Everything. Anthropic’s docs say it’s for isolated containers and VMs only

Shift+Tab moves through the modes, and the status bar names the current one.

Accept edits approves edits to files in the project folder. I like plan mode for larger projects, where you work through a whole strategy before anything changes. It isn’t needed for a small edit like one measure, but it suits planning something like a whole new Power BI project. Auto mode tries to strike a balance between reducing risk and keeping interruptions to a minimum.

Bypass permissions stays out of reach unless you allow it

Bypass permissions isn’t in the Shift+Tab cycle unless you start Claude Code in a way that allows it. Starting with --allow-dangerously-skip-permissions adds it to the cycle without turning it on. Starting with --permission-mode bypassPermissions puts the session in it from the start, and that’s the option Claude Code suggested when I typed claude. Without either one, the mode isn’t there to pick.

Administrators can block it in managed settings, by setting disableBypassPermissionsMode under permissions to "disable". I expect most enterprises will disable it in most cases. If it’s available to you, make a conscious decision about whether it’s right for your work.

You’re responsible for what your AI does and the impact it has. The more tools you connect it to and the more permissions you give it, the more destructive its actions can be.

7:46Folder context and the context window

How did the agent know where to add the measure? The prompt named the Sales table, and Claude Code was working in the project folder: the folder open in VS Code, which was also the terminal’s location when I typed claude.

Claude Code works in the folder where you start it. Anthropic’s permissions page says it has access to the files in that directory by default, and reading them doesn’t need your approval. It doesn’t read everything up front, because that would fill and clutter its context window; it reads what it needs as it works. You can give it access to other folders, with --add-dir <path> when you start or /add-dir during a session, but the project folder is where it does most of its work.

The context window is the model’s working memory for the session. Everything it reads and writes takes up room in it, and /context shows what’s using that room. Managing the context window is a topic of its own, outside this guide.

8:57@ tags: point it at one file

Even with the whole project in reach, you can give the agent a hint. Type @ and a path, and that file comes into your message. The suggestion menu lists paths from the open folder, so for a file a few folders down, type into the path, such as @Sales.SemanticModel\definition\, and pick the file from the list.

Typing into the path narrows the @ menu to the model’s TMDL files.

Anthropic’s docs say an @ file reference includes the file’s content in the conversation, up to the Read tool’s limit of 25,000 tokens by default. An @ folder reference shows the folder’s file listing, not the files’ contents. With Sales.tmdl tagged that way in a prompt that asks to update one of its measures, the agent looks in that file.

Think of it as instructions to someone who works for you. You can say “go update this measure”, or “go update this measure, and here’s where to start”. They might succeed either way, but the second makes it easier.

9:59Skills: reusable prompts you call with /

A skill is a reusable prompt: it tells Claude Code how to do a job, with the steps to take, the guidelines to follow, and anti-patterns, the things to avoid. In Claude Code, a skill is a folder with a SKILL.md file of instructions, plus any supporting files. Type / and the skill’s name to run it; Claude can also use a skill on its own when your request matches it. Typing / opens a menu of what you have, and plug-in skills show the plug-in’s name first, such as /powerbi-authoring:powerbi-report-authoring in my setup.

The status line is a fun one to try, although it’s a built-in command rather than a skill. /statusline sets up a line under your prompt that shows what you choose, such as how full your context window is or which model you’re running. It takes a description in plain English, writes the script for you, and updates your settings. Mine shows the Git branch, the session name, the model and its context size, a 10-block bar with the percentage of context used and the tokens used out of the total, the effort level, and my 5-hour and weekly usage limits with their reset times. To build one like it:

/statusline show the git branch, the session name, the model with its context window size, a 10-block bar with the percentage of context used and tokens used out of the total, the effort level, and my 5-hour and weekly usage limits with their reset times

/skills lists every skill you have. In the demo it listed 56, each with a token count. The ones on screen all come from the fabric-skills plug-in.

/skills lists every skill you have, with a token count for each.

Skills you add yourself come in two kinds:

• A project skill lives in the project’s .claude/skills/ folder. If the folder is a Git repository, the skill is part of the repository, so everyone on your team can use the same one. Switch to another project, and the skill isn’t there unless that project has its own copy.
• A personal skill lives in .claude/skills/ in your home folder and works in every project on your machine, but it isn’t on anyone else’s.

Claude Code loads every skill’s description into each session, so it knows what’s available, and a skill’s full instructions only when the skill is used. The descriptions still add up, which is why not every skill should be a personal one. With lots of skills, your context window gets crowded and harder to manage, and you get less effective results, more slowly. Don’t turn on skills for the sake of having them.

12:04Writing your own skills and vetting third-party ones

Some skills ship with Claude Code, plug-ins add more, and you can write your own. I don’t really write mine: I have Claude write them, with the writing-skills skill from the Superpowers plug-in (/superpowers:writing-skills). When I get to the end of something I’ve worked out for the first time, I ask Claude to write a skill that captures it and documents it, anti-patterns included. Superpowers is in Anthropic’s official plug-in marketplace:

/plugin install superpowers@claude-plugins-official

Its README describes writing-skills as “Create new skills following best practices (includes testing methodology)”.

Skills are written in prose, so you can read them. Some come with reference files, such as Python scripts, that are harder to read if you’re not a coder, and you can ask Claude to review those.

I’d caution you against skills from third parties. If you use one, investigate it first: read it, understand what’s in it, and be careful of anything that refers to an external resource. Stick to skills from publishers you trust. Anthropic’s skills page adds one thing to check: a skill can grant itself broad tool access through its allowed-tools field, so review that field in any skill checked into a repository before you run Claude Code there.

Microsoft’s Skills for Fabric is the kind of trusted source I mean: Microsoft released it at Build, with a few dozen skills for Fabric development and about six for Power BI.

13:30Wrap-up and what’s next

For your first session, start in plan mode, and start with something basic. Learn the mechanics of how the harness works, and read every step it takes. As you get comfortable, give it more autonomy.

My setup might be different from yours, and I’d like to hear what yours looks like: leave a note in the video’s comments. To take the Power BI skills further, How To Use Microsoft’s Fabric Skills To Build a Power BI Report builds a full report with them. To hand Claude Code admin chores in the service, such as checking who has access or which refreshes failed, Automate Fabric and Power BI Admin with Claude Code sets it up with the Fabric CLI.