Writing effective prompts for Copilot

Use the elements of effective prompts

A prompt can be a very simple instruction, such as “Write a blog post about the benefits of remote work,” or it can be a much longer prompt including specific detail about the context and desired output, along with examples of what you want to include or create. Sometimes a simple instruction will be enough to generate what you need, but other times you will find that providing Copilot with more detail and more information will give you a better result.

In this section, you will learn the four main elements you can use to write effective prompts and how each element affects the response you get from Copilot.

The elements of effective prompting for Copilot are:

• Goal

• Context

• Source

• Expectations

You will learn how to use these elements through a series of examples using different Copilot Pro experiences. This section shows prompts and responses in Copilot on the web and Copilot in Word, to help you understand the different prompting scenarios and examples in context.

Goal

When you first write a prompt, you are likely to provide Copilot with a simple statement or instruction telling it what you want to do. Usually, this simple instruction is a goal. For example, “Write a blog post about the benefits of remote work.”

This type of prompt will produce a very general response. This can work well if you don’t have something specific in mind or if you are using Copilot to help you brainstorm ideas or understand more about a topic to help you get started, but you are not likely to get what you want from Copilot with a simple general prompt like this. The results will usually feel very generic and not likely aligned with what you had in mind.

Return to the idea that prompting Copilot is like delegating a task to an assistant. If you gave this instruction to an assistant exactly as it is typed here, what might you expect? Most likely you’d get a very general blog post on the topic, covering the most widely acknowledged benefits. If you were in a workplace where the assistant also had additional context, they might apply that; but in this case, Copilot doesn’t have anything else to work with. It only has this very general instruction with no context.

The result is a well-structured, easy-to-read, but very generic blog post.

If your request is very specific, you may get a satisfactory result from a simple prompt like this, whereas if it is a request that can be broadly interpreted in different ways, the result will be less predictable. Compare the following two examples:

Example 1: A simple prompt with a broad goal.

The prompt “Create a workout plan” is very broad. Imagine approaching your assistant and asking them to do this. There is no information here about who the plan is for or what you want to achieve. You would be unlikely to get a great result from asking them to do this with no context or expertise.

As expected, Copilot returns a very general response to the prompt. It starts by saying “it depends” and then proceeds to offer some general advice and links to resources. This is not likely to be particularly useful.

Example 2: A simple prompt with a specific goal

In this example, the prompt is also very simple, “Create a recipe for vegan gingerbread,” but the request is very specific. Gingerbread is made up of largely the same ingredients, and there are a limited number of ways to make it vegan.

In this case, the Copilot response is specific and well aligned to the instruction. This is likely to be a satisfactory output without needing any additional context or information in the prompt.

Context

One of the most impactful things you can do to improve the responses you get from Copilot is to add context to your prompt. This helps Copilot understand the “why” behind your instruction and any other detail about what else you had in mind when you came up with the task. Think about how you would delegate this task to an assistant. You would likely start by positioning the request with information about why you wanted it done and what it will be used for.

You can add context to your prompt by describing the purpose, the intended audience, or how it will be used. Providing context completely changes the result you will get from Copilot from something generic to something that more specifically aligns with your goal.

Compare the following prompt with the simple instruction used in the previous step, where you asked Copilot to write a blog post about the benefits of remote work. This prompt now describes the intended audience (“small business owners”) and goes a step further to provide even more specific detail about the mindset of the audience (“who are considering allowing their employees to work remotely”). It also describes the “why” behind the goal – “to highlight the advantages and persuade them to adopt remote work policies.”

This prompt gives Copilot a lot more to work with, and the result is very different from what you get with a general prompt without all this context. Context matters because it helps Copilot give much more tailored results. The resulting blog post is now written specifically for this audience and purpose. The heading, subheading, content, tone, and explanation of the benefits are different, and the result is much more tailored and less generic.

You can provide a lot of detail in your context or use a simple statement. In the example where we asked Copilot to create a workout plan, adding the context about who the plan is for makes a huge difference to the relevance and usefulness of the response.

Now that Copilot knows the workout plan is for a beginner looking to improve their fitness, the result is much more specific, and likely something that you could put into action, rather than the generic advice it provided from the broad prompt.

Compare the following result when you give Copilot the same basic instruction with a completely different context.

As you would expect, the response is different and aligned to the context provided.

Source

You can refine your prompt and the output by providing Copilot with source material to work with. Source material can include additional information or examples in the prompt; content already in your documents, notes, or presentations; or content from the web.

Let’s return to the blog post about remote work and develop the prompt further to include examples. Imagine again that you are briefing your assistant on the blog post you want. You have provided the goal and the context of the intended audience, but you also have in mind some specific facts or examples that you want to make sure are included. When you delegate that task, you provide that information at the same time. The prompt now includes details of two example scenarios that illustrate the benefits of remote work for small businesses.

The response from Copilot includes the requested examples in the text.

Copilot uses a large language model to generate the response, as described earlier in this chapter, which means it doesn’t understand what it is generating. Copilot is using the large language model and the training data to create content by predicting likely words and sentences. That means sometimes the response includes content that looks true but isn’t. This is known as “hallucination.” In this response, the cited Stanford University study sounds plausible, but it doesn’t actually exist. Copilot is a very effective assistant, but like working with any assistant, you should always check the work it does before relying on it. The more you use web plug-ins or other content in your prompt, the less likely you will find hallucinations in the response.

Copilot Pro allows you to use the content of your documents as a source to create new content in some of the Microsoft 365 app experiences.

When you use Copilot in Word, PowerPoint, and OneNote, you can use a simple prompt to summarize the content. In the prompt “Summarize this doc” in Copilot in Word, the phrase “this doc” instructs Copilot that the document is the source you want it to use.

Copilot works with the source provided and responds with a summary of the content in your document.

You can also write prompts to ask Copilot to rewrite or draft new content using the content in your document or notes as a source by using a phrase such as “based on this doc” in your prompt.

Copilot generates a social media post suitable for LinkedIn, using the content of your document as the source.

You can use a Word document as a source to create a presentation using Copilot for PowerPoint by including a link to the document in the prompt.

Copilot creates a presentation outline based on the source document and uses the content in the document to create the presentation content, speaker notes, and images.

You can also add source content to your prompt by copying and pasting it from the source directly into the prompt area. In this example, you have a vegan gingerbread recipe as your source, and you want to rewrite it to make it easy for children to understand.

Copilot responds with a new version of the recipe written in a different style.

Expectations

Expectations define the length, tone, or format of the output you have in mind. You can ask Copilot to write to a specified word count or to respond in a different format, such as bullet points. You can describe the intended audience or type of language used.

Just like delegating a task to an assistant, you will find that you get better results if you explain the output style and format you have in mind. Compare the following prompts:

• Summarize this doc

• Summarize this doc in three bullet points

• Write an executive summary for this doc

Each of these prompts asks Copilot to summarize the content of the document, but the intended output is different each time.

• When you ask Copilot to summarize this doc with no other explanation of the desired output, you will get a comprehensive summary, which may be written in bullet points or paragraphs, as you saw in the previous section.

• Adding the expectation that you want the output written “in three bullet points” will provide you with a much more concise summary, written in a bullet point format.

• You can express the desired output as a style, such as asking for an “executive summary.” In this case, Copilot will summarize the key points of the document focusing on the main conclusions and recommendations for a business decision-maker.

You can describe in detail your expectations from the output, including structure, content, and tone. Here is an example of a prompt that now includes the desired length (1000 words) and specific subheadings (Increased Productivity, Cost Savings, Employee Satisfaction) and asks for a concluding paragraph. It also specifies the tone that should be used (professional yet friendly).

Write a 1,000-word blog post about the benefits of remote work. This post is intended for small business owners who are considering allowing their employees to work remotely. The goal is to highlight the advantages and persuade them to adopt remote work policies. Include examples such as a tech startup that saw a 20% increase in productivity after transitioning to remote work, or a marketing agency that reduced overhead costs significantly by closing its physical office and moving to a fully remote setup. The post should have at least three subheadings (e.g., Increased Productivity, Cost Savings, and Employee Satisfaction) and a concluding paragraph that summarizes the key points. Maintain a professional yet friendly tone throughout.

The response follows the expected output requested in the prompt.

Compare the output with a prompt on the same topic that sets a very different expectation.

From this prompt, Copilot produces a very different response.

Iterate and refine

In this section, you have learned the main components that make up a prompt—goals, context, source, and expectation—and how to use them to get better results. However, it is unlikely that you will ever get a perfect result the first time from a single prompt. It is important to understand that generative AI will not always correctly understand your intent the first time.

Think again about delegating a task to an assistant. You may need to go back and provide additional information, context, source material, or expectations to get the results you were hoping for. More likely, you would engage in a conversation either at the start or throughout the process with the assistant to clarify, refine, and provide additional detail.

When working with Copilot, think about the interaction as a conversation, as you might have with an assistant, rather than only as a single prompt and response. Developing your skills in this area and learning how to iterate and refine your prompts to get better responses is the most important factor in your success using Copilot.

Returning to the example of using Copilot to create a workout plan, when we first asked Copilot to “create a workout plan for a beginner who is looking to improve their fitness and has not had a regular exercise or workout routine before,” we got a response that suggested a simple six-week workout plan.

However, it was a general response with very little detail. Perhaps you had in mind that you wanted a four-week program to start with, or you need to provide more details about the exercises to help beginners get started. You can refine the results by responding with another prompt.

Copilot takes your feedback and refines the initial response to provide the changes and additional details you requested.

You can continue the conversation in this style until you get the result you want, iterating to ask for more changes or for additional information to be added.

Additional tips to improve prompting

Goals, context, source, and expectations are the core elements of writing effective prompts. You can improve your prompting even further with the additional tips covered in this section.

Be specific

In the previous section, you learned how to make your prompt more specific by adding context, examples, and expectations about the output. Providing detailed and clear instructions helps guide Copilot to give you a better result. This is true even for a simple prompt. Compare the following two examples. Even without adding context or examples, you will get better results by being more specific about what you want Copilot to do.

• Tell me about cats

• Describe the different breeds of domestic cats and their unique characteristics

Tell Copilot when you are starting a new conversation

Copilot expects you to have an ongoing conversation where you iterate and refine your prompts, so when you are ready to start a new topic or conversation, you should tell Copilot. That effectively resets the conversation and avoids confusion.

In the Copilot experiences on web, mobile, and Windows, there is a button you can select next to the prompt area that restarts the conversation.

In the Copilot experiences in the Microsoft 365 applications, look for the option to click a button to change topic, just above the prompt area.

Use quotation marks

You can use quotation marks in your prompt to provide greater clarity and precision. For example, if you ask Copilot to describe “machine learning,” providing quotation marks around the term will focus the output precisely on defining this term.

The same prompt without quotation marks will vary in the focus of the response, giving you a more general answer.

Quotation marks can also help when you want Copilot to use an exact phrase or quote rather than paraphrasing it. They can also help to provide clarity on where the quote is in your prompt, rather than the words running together in an ambiguous way.

Quotation marks are also useful when you want to refer to a title in your prompt:

Use tags or labels

You can help make your prompt clearer by using tags to identify different parts of your prompt or describe the desired output. In this example, the prompt asking Copilot to write about the benefits of remote work also includes tags describing the structure and what to include in each section.

The response follows the requested structure and description of the content.