The four building blocks of a good Copilot prompt

In briefWhat this article is about

The same moment turns up in nearly every training session. Someone types a question into Copilot, reads the answer, shrugs and says it is not all that impressive. When we then look at the question together, nine times out of ten there was too little in it to produce anything usable. Not badly worded, just empty.

In its own documentation Microsoft describes four parts a Copilot prompt can be made of: the goal, the context, the source and the expectations. That list is short enough to remember and concrete enough to repair a failed question with. This article walks through all four, shows what happens when one is missing, writes the same question four ways, and ends with the things no wording can fix.

01Four parts, and what happens when one is missing

The four parts are chosen to cover what a colleague would know and a language model would not. Ask a junior colleague to draft something and you naturally say what it is about, who it is for, which documents belong to it and roughly how long it should be. With Copilot people skip those last three, because the box looks like a search bar and nobody talks to a search bar.

Microsoft is clear that only the goal is required. The other three are optional and improve the answer. That is accurate, but it hides where the gain sits: the goal decides whether you get an answer, the other three decide whether you can use it. In practice that is the difference between a text you still have to rewrite and a text you change two sentences in.

The value of the list is not in memorising it either. It is in recognition. When an answer turns out useless, run through the four and see which one was absent. Usually it is exactly one, and the improved prompt is half a sentence longer.

02Goal: start with the verb

A goal is a verb with an object. Summarise. Draft. Compare. Check. Convert. What people actually type is often not a goal but a subject: "the annual report", "the customer survey", "this quotation". Copilot then invents a goal of its own, and rarely the one you meant.

Two things go wrong at this stage. The first is several goals in one prompt. "Summarise this report, pull out the risks and write an email about it to the board" is three requests. Copilot does all three halfway. Give them one after another, so you can steer in between and use the answer to the first as the starting point for the second.

The second is the request that is not a request. "What do you think of this text?" produces a polite piece of writing with nothing in it. "Name three places where this text is unclear to someone who does not know the project" produces three places. The narrower the goal, the more usable the answer, even though starting narrow feels like giving something away.

03Context: what a colleague knows and the model does not

Context is the part people skip because it feels obvious. You know the letter is going to a local authority, that the reader is not technical and that the file has been turned down twice before. Copilot does not know any of that unless you write it down or send it along.

Useful context answers three questions. Who is it for? What is it meant to achieve? What does that reader already know? With those three added, a text changes from a general piece into something that fits your situation. An example from a session at a local authority: the same request to summarise a council paper first produced a tidy summary, and with "for councillors who do not follow this file, who mainly want to know what it costs and what is being asked of them" added, it produced a summary the officer involved forwarded without changes.

There is a limit. Context is not a backstory. Three sentences almost always do what a paragraph would do, and too much background works against you: the more there is in the prompt, the more there is to dilute it. Write down what changes the answer, not what completes the picture.

The place where context pays off most is tone. People who routinely say what kind of reader a piece is for spend far less time afterwards reworking the wording. It is also the part that differs most by role, and the reason our Copilot programmes run differently per job family: a management assistant has a different standing context than someone in finance.

04Source: the part most often left out

Of the four this is the one with the biggest jump, and at the same time the one least used. Without a source, Copilot answers about your field. With a source, it answers about your work.

You point at material with a forward slash. Type one in the prompt box and a list appears of files, emails and meetings you have access to, from which you pick what belongs to the request. Dragging a file in or adding it with the plus button works just as well. Anyone who has seen this once uses it daily; anyone who never has pastes text in by hand and runs into length limits.

This is also where the sharpest difference between the free and the paid variant sits. Pointing at a file you supply yourself works without a paid licence. Having Copilot find the answer on its own, without you saying where it sits, does not: that requires the layer allowed to look through your email, meetings, documents and sites. What exactly is and is not included in the free variant is set out in Copilot Chat: what do you get free, and what not.

Two more things about sources. Several sources at once is possible, and it is often the moment people see the value: putting two versions of a contract side by side and having the differences listed is an afternoon of work that is done in a minute. And put the reference near the end. Microsoft notes that later parts of a prompt carry more weight, which shows most in long prompts with many instructions.

05Expectations: the form you want it in

The fourth part is the cheapest and returns time immediately. Say how long it should be, in what form, in what tone, and what should stay out. Five words at the end of a prompt save a round of rewriting.

• Length. "In at most two hundred words" or "in five lines" works better than "short". Short is an opinion, two hundred words is a number.
• Form. A table, a numbered list, an email with a subject line, a running paragraph. Ask for nothing and you get prose with headings, which is rarely what you needed.
• Tone. Businesslike, informal, formal, explanatory for someone with no background. This is the setting that cuts rewriting time hardest.
• Language. If you work in English with a document in another language attached, say which language the answer should be in. Otherwise Copilot regularly follows the language of the source.
• What should stay out. Microsoft advises phrasing instructions positively: say what should happen rather than what must not. "Stay with the facts in the document" works better than "do not invent anything".

That last point surprises people, because in human terms a prohibition feels clearer. With a language model it works the other way round: a prohibition names the thing you wanted rid of and therefore keeps it in view. A positively phrased instruction describes the behaviour you do want, and that is easier to follow.

06The same question, written four ways

In the abstract this stays vague. Below is the same underlying need four times, written with one more building block each time. The situation: you have had a meeting about a project that is running late, and you have to bring the client up to date.

The jump sits between the second and the third row. That is where drafting a text ends and processing your own material begins. It is also the reason we do not work with sample prompts in training but with material the participants bring themselves: anyone who makes that jump once with their own file does not need convincing afterwards.

07Four things Microsoft adds itself

Alongside the four building blocks the documentation contains a few pointers that matter in practice and that few people know.

• Order counts. What sits at the end of a prompt carries more weight than what sits at the start. So put the source and the most important instruction at the end, not in the middle of a long paragraph.
• Expect several rounds. Microsoft says so itself: the first answer is rarely the final one. Following up on what is there is cheaper than rewriting the prompt from scratch. "Make the second paragraph more concrete" is a complete instruction.
• Start a new topic when you change subject. Copilot carries the conversation so far along with it. Keep asking about something else in the same window and the previous question colours the new answer. Say that it is a new topic, or start a new conversation.
• Use quotation marks. If you want one particular sentence rewritten, put it in quotation marks. That removes the discussion about what exactly was meant.

The documentation also says, and we repeat it in every session, that the result is something you check. A language model predicts words and can write something untrue with complete conviction. That is not a reason to avoid it, it is a reason to use it for work whose result you are able to judge.

08What no prompt can fix

There are four situations in which tinkering with the question is pointless. Recognising them saves an afternoon of frustration.

1. The source is out of reach. If the answer sits in a folder you have no rights to, or the connection concerned has been switched off by the administrator, nothing comes back. Copilot does not add access you did not already have. That is not a wording problem but an administration question.
2. The work belongs to a different tool. A calculation with hard constraints belongs in Excel, not in a conversation. A recurring process belongs in an automated flow or an agent, not in a prompt someone retypes every week.
3. The question involves something confidential that does not belong there. A commercially protected environment does not mean everything may go into it. What you do with special categories of personal data and third-party files is an agreement the organisation makes, not something each employee works out alone.
4. There is no right answer. Ask for a judgement the organisation has not formed a position on and you get a plausible-sounding opinion. That reads well and is worthless.

The first two are the ones we meet most often, and both can be solved without any training. The third calls for an agreement on paper. That belongs to the rollout rather than to the skill, and the article on why Copilot rollouts stall goes into it further.

09From a single prompt to a shared library

A prompt that works well is work you do once. Retyping it every month is a waste, and keeping it to yourself even more so. Microsoft has set up a place for this: the Prompt Gallery, reachable from Copilot and Copilot Chat.

Three kinds of prompt sit there side by side. Under suggested are the prompts Microsoft has written, each with a note on how to adapt it to your own situation. Under your prompts are the ones you have saved; those are visible only to you. And under teams are the prompts someone has shared with a Teams team, which is how a department builds up a library of its own in practice. Saving is done with the bookmark icon on a prompt you have just run.

For anyone looking at this from a distance there is something else that is useful: an administrator can export which prompts have been saved, liked and shared, per user and per team. That is not meant as a monitoring tool but as a view of what is actually being used, and it is more informative than a survey. The prompts themselves are stored inside the organisation’s own environment, encrypted, in the same storage layer as the rest of Microsoft 365.

We often see organisations take the reverse order here: first fill a library with fifty examples, then hope people use them. That rarely works. What does work is a team that spends three weeks working with it, and shares the prompts that stick. Five good prompts that come out of the work itself return more than fifty that come from outside.

In closingA level-headed conclusion

The four building blocks are not a trick and not a method with a name. They are a ten-second checklist: is there a verb in it, does it say who it is for, does it say where the answer should come from, and does it say what it should look like. Anyone with those four habits gets something different out of the same tool than the colleague who does not, without a euro extra being spent.

What it does not solve is the question of which work lends itself to this in the first place. That is not a matter of wording but of looking at your own week: which pieces do you write again and again, which documents do you read through to extract three lines, which emails cost more time than they are worth. That conversation is the real beginning, and it calls not for a course but for an afternoon with your own material in front of you.

If you would rather do that with guidance, our prompt engineering training is built for it: from the four building blocks through frameworks, supplying examples, chaining prompts and building a library of your own, model-independent so it applies to the tools you use alongside Copilot. If your team works mainly in Copilot Chat, the Copilot Chat training sits closer to daily practice. For anyone who would rather read than book, the complete guide to Microsoft 365 Copilot covers what is possible per application.

SourcesVerified on 12 September 2026

• support.microsoft.com – Learn about Copilot prompts, the four parts goal, context, source and expectations, the note that only a clear goal is required, and the examples of a simple and a more detailed prompt
• support.microsoft.com – Write a great prompt in Microsoft 365 Copilot, referencing sources with the forward slash, using quotation marks and starting a new topic
• support.microsoft.com – Get better Copilot responses with great prompting, the order within a prompt, phrasing instructions positively, working in several rounds and the call to always review the result
• learn.microsoft.com – Understand Prompt Gallery in Copilot, the three kinds of prompt side by side, the export of saved and shared prompts by an administrator, and storage inside the organisation’s own environment
• support.microsoft.com – Save your favorite prompts to use again, the bookmark icon and the note that saved prompts are visible only to you
• support.microsoft.com – Sharing prompts with your team, sharing to a Teams team and where shared prompts appear for colleagues