A Beginner’s Guide to Using AI Builder Prompts in Copilot Studio
Building an agent in Copilot Studio is simple. But making it useful, that’s where many people get stuck. Out of the box, your agent can answer questions, but it doesn’t always know how to do things in a way that matches your business needs. This is where AI Builder prompts come in.
Prompts are like smart instructions you give your agent. With just a few lines, you can teach it to summarize an email into three key points, classify customer feedback as positive or negative, or even pull totals from an uploaded invoice.
In this guide, we’ll go step by step to show you how to build prompts, test them, and use them to create an agent that feels sharp, reliable, and ready for real-world tasks.
What Are Prompts in Copilot Studio?
Think of prompts as custom actions you drop into your agent. Instead of relying on generic AI replies, you tell the system exactly what to do, what input to expect, and how the output should look.
Here’s why that matters:
- With a Summarize prompt, your agent can turn pages of text into a short digest.
- With an Extract prompt, it can pick out names, dates, or order numbers automatically.
- With a Classify prompt, it can label messages so your workflows know where to send them next.
In short, prompts move your Copilot Studio agent from being a “chat responder” to being a task performer. And once you learn the basics, you can reuse them across multiple topics to keep your agent consistent and efficient.
Setting Up Your First Prompt
Before you start, make sure you have:
- A Copilot Studio agent was created.
- Access to the designer (Topics and Tools).
- A clear task in mind for your prompt.
Now, follow these steps:
- Open Copilot Studio and select your agent.
- Go to Topics and open a topic where you want to add the action.
- Click the “+” icon and choose Add node.
- From the menu, select Add a tool → Create a prompt.
- A new window opens where you can define your prompt.
At this stage, you’ll name the prompt, write instructions, and add inputs.
Available input options are:
- Text: Pass a string of text.
- Image or document: Attach a file for the model to analyze.
- Power Fx: Supply a calculated value.
Step-by-Step Example: Sentiment Analysis and Reply
Let’s build a prompt that reads a customer message, detects sentiment, and drafts a reply.
- Inside the new prompt, give it a name like “Customer Sentiment Handler.”
In the Instructions box, write:
You will read the customer’s message in {{Details}}.
1) Identify the sentiment: Positive, Neutral, or Negative.
2) Create a short reply that matches the tone:
– If Negative: Apologize and say a support ticket has been created.
– If Positive or Neutral: Reply politely without apology.
3) Output in JSON format:
{
“sentiment”: “Positive | Neutral | Negative”,
“reply”: “string”,
“ticket”: “Created | NotNeeded”
}
- Add an Input called Details (type: Text). This will receive the customer’s message from your topic.
- Use the Test panel with a sample message (e.g., “I’m really upset about my order delay”). The model should return JSON with “Negative” and a polite reply.
- Save the prompt.
- Back in your topic, map the user’s message into the Details input, and capture the JSON output into variables you can use later.
Working With Documents and Images
Prompts can also process files. For example, if you want to extract totals from an invoice:
- In the prompt editor, add a new Input with type Image or document.
Write instructions like:
From the provided invoice file {{InvoiceDoc}}, extract:
{
“invoiceNumber”: “string”,
“totalAmount”: “currency”,
“dueDate”: “date”
}
- During a chat, ask the user to upload the invoice. The file is passed to InvoiceDoc, and the model extracts the key fields.
- Save the output in your topic variables to update records or trigger workflows.
Practical Prompt Templates You Can Reuse
Summarize Content
Read {{Content}} and return:
{
“summary”: “3 sentences”,
“highlights”: [“point1”, “point2”, “point3”]
}
Extract Fields
From {{Message}}, extract:
{
“customerName”: “string or null”,
“orderId”: “string or null”,
“issueType”: “billing | shipping | product | other”
}
Classify Feedback
Read {{Feedback}} and return:
{
“category”: “bug | feature | praise | complaint”,
“reason”: “short explanation”
}
Best Practices for Strong Prompts
To keep your prompts reliable and easy to maintain:
- Be explicit about output: Always define a JSON schema.
- Keep instructions short: Clarity improves accuracy.
- Use examples: Show the model exactly what you expect.
- Test with real data: Use actual messages or documents.
- Reuse prompts as tools: Store common prompts under Tools for consistency.
Conclusion
AI Builder prompts transform Copilot Studio agents from generic bots into purposeful helpers. By defining clear instructions, using structured outputs, and testing with real inputs, you can create agents that summarize, extract, and respond intelligently. Start with one prompt, refine it, and expand into more advanced cases as your needs grow.