Resources & insights

Are you asking Microsoft Copilot the right questions? Top tips. 

AI-powered solutions like ChatGPT, Microsoft Copilot and Copilot for M365 have changed the way we search for information, analyse data and create content.  

Think of it as your AI-powered, always-available workplace assistant, sitting quietly in the Microsoft apps you use daily, just waiting to be called upon. You can use Copilot for M365 to complete the tasks that usually take up your time, like summarising a Teams meeting or drafting an email to a client. What’s more, it’s easy to use. You simply tell Copilot what to do in your own words. 

However, the key to getting the best results out of these new tools is being able to craft the right question. When you’re clear about what you want and convey it to Copilot precisely and with context, you’re more likely to get an accurate result the first time. However, sometimes it’s not simply as easy as asking the right question; the way you ask it can be just as important and sometimes, surprising questions give surprising answers. Crafting these questions has become a skill in itself, known as prompt engineering. 

In this article, we’ll explore prompt engineering in more depth. Firstly, we’ll look at how prompts work. Then, we’ll share some prompt engineering tips, so you can use Copilot to save time and deliver better work. Let’s get started. 

How prompts work 

To craft the most effective prompts, it’s useful to know how large language models work.  

One thing to remember is that LLMs don’t actually know anything or store information. Rather, they’re highly adept at sentence construction. LLMs are trained on vast amounts of data. For example, OpenAI (creators of ChatGPT and the ChatGPT-4 model used by Microsoft Copilot) scrapes the internet for publicly available text data to train its LLM. This text data is then processed through an AI engine, which reads the data and identifies patterns in how words relate to each other in certain contexts. By identifying these patterns on a vast scale, the LLM can then reasonably predict which words should follow another. 

Ask the LLM a question, and it pieces together an answer word by word, with each word being the most natural follow-on. Sometimes, however, the words and phrases it generates will not make sense, or at least not be as accurate as they could be. That’s where prompt engineering comes in. 

A prompt is all about coaxing the right words from the LLM. The more effective your prompt, the more likely you are to get the insight you desire. On the other hand, a poorly constructed prompt could generate an inaccurate response, leading to bad outcomes if you use it in your work. 

Constructing your prompt 

At its most basic, a prompt is nothing more than a command or question. It could be ‘Create a business plan for my e-commerce startup’ or ‘Tell me what Tom Cruise movie I should watch next’. However, if you are more specific with your prompt, you will get better results (or at least a result that resembles what you want more closely). 

While the first versions of the LLM could only support relatively small amounts of data, newer models can support data the length of the average book, so don’t be afraid to give highly-detailed instructions. 

Firstly, give as much information about what you want as you can. If you only ask, ‘Write an investment memorandum’, you’re probably not going to get anything useable. Instead, say, ‘Create a new investment memorandum based on the heading in other documents. Include key information XYZ about this client. Also include our track record in the sector.’ Copilot will fulfil all those commands, and you’re likely to get something valuable.  

Think of the information you give in your prompt as a constraint. If you asked for 20 reason why someone should invest, they’ll probably be quite tenuous or repeats of previous ideas by the time you get to the last few. But limiting it gives focus to the result. 

Next, provide as much context as you can. Returning to our investment example, begin the prompt with something like, ‘You are the editor of The Economist. You’re highly knowledgable about finance and investment and you write with clarity and purpose.’ That prompt gets Copilot to pretend they’re that expert, and the document it creates will likely contain more of that specialist expertise. In contrast with ChatGPT, Copilot will need slightly less hinting because it’s already grounded in your data. 

You could also give examples to add even more context. For instance, you could ask Copilot to provide you with ‘10 effective opening lines for my investment document, similar to…’, then inserting an example that has worked in the past. 

Another best practice is to be specific about the form of the output you want. For example, you could ask Copilot to produce your investment document in HTML format so you can paste it straight into your website’s content management system.  

Essentially, nothing is off-limits when it comes to prompt engineering. You’re unlikely to run out of space in the prompt box. Ask for what you want clearly, and you’re likely to get it. Then, use further prompts to drill down until you have precisely what you want. 

Prompt engineering best practices 

While the effectiveness of your prompt – how you ask the question – is the primary factor in the response you receive, there are more things you can do to turbocharge your prompt engineering 

• Go deeper – Try to learn more about Copilot as you use it. For example, while you could ask Copilot for an Excel formula that you could paste into a document, you could also ask for a detailed explanation of what the formula does and why it’s the right tool for this particular job 
• Feed in your data – With Copilot for M365 you can securely add your organisation’s data as part of the prompt. Tell the LLM which data sources to refer to so you can be sure it’s referencing the correct information 
• Keep your data organised – If you’re a Copilot for M365 user, ensure your organisation’s data is correctly stored and labelled, with the access permissions set. This ensures the LLM accesses the right data when you prompt it 
• Start a prompt library – When you discover a prompt that gives you exactly what you asked for, saving you time and effort, copy it into a document so you can use it again. Better yet, create a shared, organisation-wide prompt library so your team members can share their knowledge and time-saving prompts 
• Experiment – Prompt engineering is a skill that requires testing, experimentation and iteration. Always try to go one level deeper with one extra prompt. The more you use Copilot, the better you’ll get at controlling it so you get the results you need. Be curious, and remember, you can’t break it! 

Prompt engineering in Microsoft 365 Copilot  

As mentioned in the section above, Copilot for Microsoft 365 gives Copilot an extra dimension by embedding It into your M365 apps like Outlook, Excel, Teams, and Word. With a few clicks, you can harness the power of Copilot to create content or perform analysis around your organisation’s Microsoft files. When used correctly, it’s a massive time-saver can can change the way your organisation works for the better. 

To get the most value from this groundbreaking capability, follow the tips and best practices mentioned above. Be clear in your instructions and provide as much context as possible. Be specific about how you want the output presented. Give follow-up prompts if necessary. 

Let’s look at a couple of examples. Imagine you have just finished a long Teams meeting and you would like to see a concise summary and get the action points. Simply ask Copilot for M365, ‘Please summarise the Teams finance department meeting from March 8^th. Give me the 5 most important points discussed and tell me any action points that relate to me. Please use straightforward language.’ Copilot for M365 will then analyse the transcript from that meeting and give you everything you need. 

Sometimes, however, a simple prompt is all it takes. For example, if you’re wondering why you declined an investment opportunity in a company and don’t have the time to go through all your emails and documents, simply prompt Copilot for M365 with, ’Why didn’t we invest in Company Y last time around?’. Copilot will search your email history and give you the answer with all the context you need. You can then drill into that data to get the real answer. 

It’s worth remembering that the effectiveness of Copilot for M365 in these cases depends on how it accesses your data. To get the best outcomes, ensure your data is correctly classified. 

Book your Copilot Readiness Diagnostic  

ChatGPT led the way by showcasing the benefits of AI. However, Microsoft has taken the technology and made it enterprise-friendly. Copilot for Microsoft 365 takes it to the next level by integrating your organisation’s data and methodology.  

Doherty Associates can help you adopt Microsoft Copilot into your business. This usually starts with a free 1-hour Copilot Readiness Diagnostic call where we will cover: 

• Assessing Readiness: understand how prepared your business is to deploy Copilot. 
• Mitigating Security Risks:  understand the potential risks and how to mitigate them. 
• Building a Business Case: discuss your reasons for adopting Copilot and how to achieve ROI. 
• Exploring Possibilities: discuss pain points and opportunities to enhance your operations. 
• Planning Deployment: discuss key user groups and use cases specific to your industry. 

Get in touch today to book your call –  Doherty – Contact page 

Back to resources & insights