Writing High-Quality Instructions for Microsoft Copilot Studio Agents

When building intelligent agents in Microsoft Copilot Studio, instructions act as the central compass for your AI orchestrator. They define how your agent reasons, which tools or knowledge sources to call, how to fill tool inputs, and how to format responses for end-users.

Without clear, structured instructions, even a well-configured agent can struggle with intent recognition, tool selection, and output consistency. Drawing directly from official Microsoft guidance, let’s explore how to author instructions that make your Copilot Studio agents sharp, reliable, and production-ready.

What Do Agent Instructions Actually Do?

Before drafting your text, it helps to understand what the orchestration layer relies on instructions for:

1. Resource Routing: Deciding whether to call a tool, knowledge source, topic, or secondary agent to handle a user query or trigger.
2. Parameter Population: Filling required inputs for tools based on conversational context.
3. Response Generation: Structuring the final output to match your exact business requirements.

Note: An agent cannot act on instructions for tools or knowledge sources it doesn’t already have. Ground your instructions strictly in capabilities you have explicitly configured.

1. Master the Slash (/) Syntax

When writing or editing instructions on your agent’s Overview page, avoid typing raw text names manually when referencing assets. Instead, use the forward slash (/) key.

Typing / opens an inline selector allowing you to reference specific resources directly:

• Tools (e.g., / "Get_OrderStatus")
• Topics
• Agents
• Variables
• Power Fx expressions

Crucial Rule: When naming a tool or object in your instructions, use the exact same name and construct it with the / selector. Slight naming discrepancies can confuse the orchestrator and degrade performance.

2. Use Action-Oriented Language

The LLM orchestrator performs best when instructions use clear, imperative verbs. Structure your prompts using specific categories of vocabulary:

• Data Retrieval: Use verbs like retrieve, get, use, analyze, extract, or provide.
• Processing & Filtering: Use terms like from, include, exclude, compare, and identify.
• Conditions: Use conditional triggers like when, if, ensure, and compare.

Example Pattern:

“Retrieve customer shipping data using / "Get_Customer_Shipping" and filter records to extract items delayed by more than 3 days.”

3. Ground and Limit Tool Usage

To prevent the agent from hallucinating or invoking incorrect actions, write instructions that place healthy boundaries on tool parameters and sequencing:

• Specify Execution Paths: Tell the agent when to execute actions sequentially. (e.g., “Only send an email update after checking / "Knowledge_FAQ" for context”).
• Limit Parameters: Explicitly state what limitations apply to tool payloads (e.g., “Only query records associated with the active user account ID”).

4. Format and Tone Guidelines

By default, Copilot Studio agents communicate politely and professionally—you do not need to waste instruction space dictating generic politeness. Reserve tone and formatting rules for specific business requirements:

• Enforce Structures: If your users expect tables or lists, make it explicit: “Always return order status results in a clean markdown table format.”
• Avoid Ambiguity: Steer clear of vague phrases like “typing box” or “chat window,” which can cause unpredictable model behavior.

5. What to Avoid: System Citations Warning

⚠️ Important: Never attempt to modify, override, or interfere with system-defined citation formats or behaviors.

• Avoid terms like “citation” or “reference” in instructions designed to alter how sources are displayed.
• Suppressing or changing citations can cause the orchestrator to fail to recognize valid results, treat them as generic model knowledge, or drop them entirely.

Summary Checklist for Your Next Agent Build

1. Pre-configure your components: Add all necessary knowledge sources, topics, and connectors before writing instructions.
2. Use the / picker: Never guess tool or variable names; pull them directly via the slash menu.
3. Be concise and action-driven: Use direct verbs and define conditional workflows clearly.
4. Test in the canvas: Use the test pane to inspect how the orchestrator builds its execution plan, then tweak your instructions iteratively.