Building Custom Prompts in AI Builder

Anyone who regularly generates text with AI in Power Automate knows the problem: the same instruction to the language model gets typed in again for every flow, slightly reworded each time, never quite consistent. AI Builder offers a cleaner solution for this: custom prompts that you build, test, and save once in the prompt generator, and that are then available as a ready-made action in any number of flows, Power Apps, and even in Copilot Studio.

This article shows how to create a custom prompt in the AI hub, what role input variables play, and how to then build the finished prompt into a cloud flow as a reusable action, including a way to keep human control over the output.

What is a custom prompt in AI Builder?

In AI Builder, a custom prompt is a saved, named instruction to a language model that runs on language models operated by Azure OpenAI Service. According to the documentation Create a prompt, this feature is limited to certain regions and may be subject to usage restrictions or capacity limitations. The key difference from a one-off prompt written directly in a flow: a custom prompt gets a name, fixed input variables, and is managed centrally in the AI hub. You build it once, test it there with sample values, and afterward you can call it as a ready-made action from any flow, Power App, or Copilot agent, without reinventing the wording every time.

Creating a custom prompt in the AI hub

The prompt generator can be opened from three environments: Power Apps, Power Automate, or Copilot Studio. In Power Automate, you sign in at make.powerautomate.com and select AI hub > Prompts > Create custom prompts in the menu on the left. After that, you proceed in three steps:

• Give it a name: Click the auto-generated name in the top left of the prompt generator and replace it with something descriptive, for example "Summarize invoice description".
• Write the instruction: You write your prompt in the text field, or start from a template in the prompt library.
• Insert an input: You type "/" or select Add content to insert, in the Input section, a field of type Text or Image or document. These inputs later receive dynamic values from your flow.

Instruction and context: the two building blocks of a good prompt

According to Microsoft, a prompt fundamentally consists of two parts. The instruction is the first part and clearly describes what the model should do, for example "Summarize this email in three bullet points". The context then provides the information the model needs for a suitable answer, for example noting that the email contains customer feedback from last week. A good prompt is also clear and concise, specific enough to steer the model in the right direction, and provides enough context so the answer is also a good content fit. If the generated answer is too long or drifts off topic, it's almost always worth refining exactly these four qualities rather than rewriting the whole instruction.

Input variables: turning static text into a dynamic building block

It's the input variables that turn a rigid block of text into a genuine, reusable action. For each input, you enter a sample value in the prompt generator so you can try out the prompt directly via Test before it runs anywhere in production. A prompt can use several inputs at once, for example one text input for the content and a second for the desired tone of the answer. Once you're happy with the result, you select Save, and from that moment on the prompt is available as a standalone building block.

Using the prompt as an action in a flow

To build your saved prompt into a cloud flow, search the actions panel for "GPT" and select the matching action. Good to know: for a long time the action was called Create text with GPT using a prompt, and since May 2025 Microsoft has renamed it, according to Use a custom prompt in flow in the English-language documentation, Run a prompt; so depending on the update status of your environment, you may still see the old name or already the new one. Either way, the process works the same:

• In the Name field, you select your previously saved prompt from the dropdown menu.
• For each input defined in the prompt, a separate field appears below, which you fill with dynamic content from previous steps, for example the subject line of an incoming email.
• The action then creates a flow variable called Text, which holds the generated answer and can be reused in any subsequent step, for example to post it as a message in a Teams channel.

Alternatively, you can also start a brand-new prompt directly from the flow designer: in the same dropdown menu, you select New custom prompt, which opens the same prompt generator you otherwise reach via the AI hub. That's handy for one-off special cases, but for reusable building blocks the route via the AI hub stays clearer, because the prompt remains discoverable there for other flows and apps too.

Keeping human control over the AI output

A prompt always delivers an answer, but not every answer should proceed unchecked, for instance before a text is sent to customers. That's exactly what an approval step after the prompt action is for: the action Start and wait for an approval on text hands the generated text to a designated person, who can approve, reject, or still edit it before approval in the Power Automate approvals menu. A subsequent condition that checks the result for "Approve" then lets the flow itself decide whether the text is automatically processed further afterward. That way, the AI stays the digital assistant that delivers the first draft, while you retain control at the decisive point.

Keeping an eye on regional availability and limits

Before you plan to use custom prompts in production, it's worth a look at the underlying conditions. Prompts run in the region where your Power Platform environment is hosted, and the exact model availability per region is documented separately, according to Feature availability by region. Microsoft also explicitly points out that the feature may be subject to usage restrictions or capacity limitations. Anyone running several flows with custom prompts in production should therefore clarify these points early on with their own IT department or Power Platform administrator, rather than discovering them only at the first capacity bottleneck. With NordFlux's Power Automate consulting, we help you set up custom prompts in a structured way from the start, so that a one-off block of text becomes a reliable, reusable action for your entire team.

Frequently asked questions

Where do I find the prompt generator in Power Automate?

You reach it via AI hub > Prompts > Create custom prompts in the left-hand navigation menu of make.powerautomate.com. From there you can start a new prompt, choose a template from the prompt library, or edit prompts you've already saved.

How does a custom prompt differ from a one-off prompt written directly in a flow?

A custom prompt is named, saved centrally in the AI hub, and can then be reused across multiple flows, Power Apps, and Copilot agents. A prompt that is only written directly in a single flow action, on the other hand, stays tied to that one flow and has to be rewritten from scratch in every additional flow if needed.

How many input variables can a custom prompt have?

The documentation doesn't state a fixed upper limit; in practice you define as many text or image-and-document inputs as your use case needs. For each input, you should enter a sample value when creating it, so the prompt can be meaningfully tested before it's used in production.

Why is the action named differently in some flows than in tutorials I find online?

Microsoft renamed the action Create text with GPT using a prompt in the documentation, since May 2025, to Run a prompt. Depending on when your environment was last updated, your interface may still show the older name, without anything changing about the actual functionality.

Can I still have the output of a custom prompt checked before it's sent?

Yes. After the prompt action, add the action Start and wait for an approval on text, and link it to the following steps with a condition. That way the generated text only proceeds automatically once a designated person has approved it.