Microsoft 365 Copilot Agents: Prompting and Instruction Tips

Building AI is now part of Microsoft 365 Copilot

Microsoft 365 Copilot now lets you build apps, automate workflows, and publish agents. Earlier in 2025, the Frontier program introduced two Copilot agents, Researcher and Analyst, to help users with multi-step research and data analysis inside Copilot.

Today, Microsoft is extending that idea to creating solutions, not just answering questions, so everyday users can move from a prompt to a working app, a running workflow, or a reusable agent inside the same secure Microsoft 365 surface.

With App Builder, describe what you need, preview it, and refine in chat. Copilot can pull from your Microsoft 365 content to create dashboards, charts, and lists, and it can store new data in Microsoft Lists so your team can share and manage it easily. The point is to give a low-code solution that helps business users make a useful app in minutes.

Workflows turns a written request into a flow across Outlook, Teams, SharePoint, Planner, and Approvals. You see each step as it’s built and can add or adjust steps in the same conversation. It runs on the same infrastructure as Copilot Studio’s Agent Flows, giving you proven reliability used across enterprises.

Inside the Copilot app, the lite experience of Copilot Studio lets you spin up task-specific agents that follow your instructions and use your chosen knowledge sources. You define purpose and tone, add SharePoint sites, files, or Teams data as grounding, test the behaviour in the same pane, and publish when ready. These agents run with Microsoft 365 governance and respect existing permissions, so they fit cleanly into your environment while giving teams a way to package repeatable prompts and content connections.

Improving Copilot Agents Outcomes

Your agent’s behaviour is shaped by two things: the instructions you give it when you build the agent, and the prompts people type when they use it. Getting both right and you will see clearer answers, fewer retries, and stronger task completion.

Standard instruction template to build Copilot agent

A well-structured set of instructions ensures that the agent understands its role, the tasks it should perform, and how to interact with users. According to Microsoft, a standard instruction for declarative agent should contain 3 main components:

A few things to remember for Copilot agent instructions

• Use Markdown (#) in headings for sections; use - for unordered lists and 1. for numbered lists; backticks (``) to highlight tool or system names (Jira, ServiceNow, Teams) and ** for critical instructions.
• Tell the agent what to do using precise verbs like ask, search, check, use. Define any nonstandard or unique terms once.
• Call out the exact action or knowledge source the agent should use. For example, “Search Teams chat history”, “Reference SharePoint or OneDrive internal documents”.
• Provide examples when complexity rises. Use a few varied examples to show edge cases and preferred tone.
• Add rules that prevent over-eager tool use, repetitive phrasing, and verbose answers.

From these 3 main components, the following example instructions are for an agent that can help resolve common IT issues. You can also check the example prompt given by Microsoft here.

# PURPOSE
You are the **Project Status Agent** for the PMO. Produce a weekly executive update from approved sources.

# KNOWLEDGE & TOOLS
- Primary sources: `PMO SharePoint` (Reports), `Project Tracker` list
- Optional data: `Portfolio API` (read-only milestones, risks)

# WORKFLOW
## Step 1: Gather context
- **Ask** for project name and date range.
- **Search** `PMO SharePoint` and **read** `Project Tracker` for the latest entries.
- If available, **read** `Portfolio API` to confirm milestone dates and open risks.

## Step 2: Draft the brief
- **Compose** one page with: Progress, Risks, Decisions Needed, Next Steps.
- **Cite** titles, record IDs, and links from sources.
- **Confirm** with the user; if changes are requested, revise once.

# OUTPUT RULES
- Clear, factual tone. Keep under 200 words unless asked for more.
- Include source titles and URLs or record IDs.

# EXAMPLES
User: "Status for Project Nova, last week."
Agent: Asks for exact dates → finds no recent files → asks for a link → user provides "/PMO/Reports/Nova_1007.docx" → composes brief → returns links and record IDs, then offers `.docx` and email paragraph.

Check Our White Paper GPT Integration in Microsoft Ecosystem

Good prompt for better answers

Good prompts tell Copilot agents exactly what outcome you want, the context it should use, and the way you want the answer presented. Microsoft’s guidance points out that a good prompt can include four parts: goal, context, expectations, and source.

• Goal: what you want Copilot to do.
• Context: facts, files, links, or constraints that matter.
• Expectations: the format, length, audience, or tone.
• Source: the document or repository you want Copilot to rely on first.

Learn some tips and tricks to improve your prompting. You can keep a prompt minimal; the only must-have is a clear goal. When you need more specific results, add other parts. In practice, you will often include more than a goal to get the result you want. Here's an example prompt in Copilot chat includes a goal and source:

Summarise customer complaints from the last 14 days using the ‘Customer-Feedback’ SharePoint folder.

And here's an example that includes a goal, context, and the expectations:

Create a two-slide update on the product launch risks. The audience is senior leadership. Use the ‘Launch-Risks.xlsx’ sheet ‘Open-Items’. Return a neutral, fact-based summary and three recommended actions.

Suitable document length for Microsoft Copilot

When you prompt Copilot, you can attach sources for it to use, such as files, excerpts, or examples. For example, you might ask Copilot in Work: “Draft a marketing plan outline based on the product features in this document.”

But to get the best output from Copilot, you should consider the length of what you provide. Because Copilot can read only a limited amount of content at once, an oversized document can reduce accuracy and speed.

Depending on the type of tasks you give Copilot, there are times that Copilot will focus only on the beginning of the document (up to its limits) and then ignore anything beyond that. For scenarios where the entire document context is needed for Copilot to provide answer, a couple of good rules of thumb on the length of the document you provide to Copilot:

• For summaries or whole-document Q&A, aim for no more than about 300 pages or roughly 1.5 million words.
• Questioning a document works best when it is under around 1.5 million words.
• Rewriting performs best on passages shorter than about 3,000 words.

When using Copilot, think of it as having a conversation with your friend. You wouldn't discuss a whole encyclopedia at once, right? Similarly, keep your document references concise and to the point for the best experience with Copilot.

Build Copilot Agents today

Turn a plain-English prompt into a working app, workflow, or agent in minutes. For a secure, practical blueprint on bringing GPT into Microsoft 365, read our whitepaper: GPT Integration in the Microsoft Ecosystem - your next step from idea to production.