From Prompts to Precision: How to Train Copilot Agents for Success
As more and more organizations are getting ready to roll out Copilot Agents in some form (usually starting with low-hanging fruits, such as SharePoint Agents and Agent Builder), the importance of upskilling and training is more important than ever. Successfully building Agents, not only requires good prompting skills as with Microsoft 365 Copilot. It requires instructing the Agents effectively. With a well-written and built Agents, end users would face less errors, even if their prompts are poorer. Here are some best practices of writing effective instructions:
- Be specific. Vague or ambiguous instructions can lead to misunderstandings and incorrect responses. Clearly define the tasks, context, expectations, and any specific requirements.
- Break tasks into smaller steps. Performance is better when the flow consists of smaller and more granular steps.
- Use clear syntax. Include punctuation, headings, and section separators that allow easier parsing of the instructions. Use delimiters between sections to separate instructions from examples.
- Use examples. Provide detailed examples and scenarios to illustrate the desired outcomes. The more precise the instructions, the better the agent can perform its tasks accurately and efficiently. Few-shot prompting involves providing the agent with a few solved examples before asking it to solve a new problem. Chain-of-thought techniques instruct the agent to provide a step-by-step solution.
- Use Do instead of Do Not. Tell the agent what to do; avoid telling it what not to do. Telling the agent what to do provides more specificity.
- Avoid contrasting instructions. Describe the expected behavior of the agent under valid or positive use cases only instead of giving both valid/positive and invalid/negative instructions. Such instructions create ambiguity and that can confuse the model. If you must instruct the agent with both valid and invalid use cases, clearly describe the queries that you want the agent to process and provide interaction examples to show the expected response to each query.
- Add knowledge. If you add knowledge to the agent, describe what is represented by each knowledge source in your instructions.
If you have more questions or
need help
