Join me exploring Microsoft Copilot through each letter of the alphabet. This week, Q is for Questions. 

(There’s a reference to a Destiny’s Child song that it’s taking all my willpower not to include here…)

Three questions, specifically. Three questions that, if you ask them in the right order, will get you better results from Copilot almost every time.

What everyone says about prompting 

There’s no shortage of advice out there on how to write better prompts. Be specific. Add context. Tell it your role. Tell it the format you want. All useful. All correct.

But most prompt guides start at step two.

Before you think about how to ask Copilot something, you need to ask yourself this:

Question 1: What am I actually trying to achieve?

If you jump into using any AI tool without doing this first, you’ll end up down a rabbit hole. 

I’m not up here on my high horse by the way, I’ve definitely done this. 

In O is for Outlook, I asked Copilot to help me with my “calendar madness” and sort out my car crash of a week. What I got back was a thorough, earnest deep dive into energy management frameworks and ideal working day design. Lovely in theory. Not really what I needed. 

That wasn’t Copilot’s fault. I handed it a fuzzy, undefined problem and expected a specific answer.

In N is for Notebooks, I sat in front of a blank canvas and drew, ironically, a blank. No clear sense of what I wanted to use the Notebook for meant no useful output from it.

When I’ve found myself struggling with Copilot’s output, I’ve realised that the problem started before I even opened Copilot. I hadn’t done the thinking first. 

I’ve started calling this end-result thinking. Before I type anything, I ask myself: what does “done” look like here? What’s the actual outcome I’m after?

Not: “Help me with my inbox.” But: “Which emails have I sent in the last two weeks that haven’t had a response?”

Not: “Sort out my week.” But: “I have a clash on Wednesday afternoon – what are my options for moving the less critical meeting?”

The sharper your picture of the destination, the better Copilot can help you get there. And honestly,  this isn’t just an AI skill. It’s a thinking skill. A muscle I didn’t expect to be working when using AI. 

Now before I even open Copilot, I’ll sometimes open a blank document and just… blurt.

Everything I’m thinking, in no particular order, with no attempt to organise it or punctuate it properly. Ideas that connect to each other, tangents, the “this connects to this connects to this” thread I’m trying to follow. What I want to see at the end. What I’m not sure about. All of it, just out of my head and onto the page.

Then I use that as my starting point for the prompt, or sometimes I paste the whole thing straight in and ask Copilot to help me work through it.

It’s worth saving that document too. If you want to tweak the approach later, or use something similar again, it’s all there, you’re not hunting through your AI chat history trying to reconstruct what you asked six weeks ago.

If typing it out feels like too much friction, Copilot Voice on your phone will transcribe your rambles for you. I’ve done my meeting prep out loud on the school run (see M is for Meetings) the same logic applies here. Sometimes thinking out loud is quicker than thinking in writing.

Now you’ve got your thoughts in order, time for question 2.

Question 2: Am I asking Copilot in a way it can actually work with?

Once you know what you want, the next question is: have you given Copilot enough to go on?

This is where most prompt guides get started, and the way you prompt does matter. Context, specificity, format, tone. The more you give Copilot to work with, the less it has to guess.

You don’t have to figure this out on your own though.

Copilot has a built-in Prompt Coach agent (I mentioned it briefly back in A is for Agents) and it really does help.

Tell it the outcome you’re trying to achieve, and it will help you construct a prompt that’s actually likely to get you there. You can use it as a starting point and keep refining, or ask it to give you a few variations to choose from.

I’ve used it when I know what I want but can’t quite find the words to describe it in a way that feels specific enough. It’s like having someone help you brief a colleague properly instead of just saying “can you sort this out?”

I used it recently to help me brief another Copilot agent, Researcher (yes I know, asking one AI tool to help me with another, we are living the film ‘Inception’) to come up with ways to grow attendee numbers for an in-person user group. 

It asked me a few clarifying questions and then gave me a detailed brief that I could copy and paste over to Researcher agent:

(Also – shameless plug for if you’re ever in London and want to hang out with other people who are interested in Copilot: https://mccug.co.uk/ )

Anyway – so now you know what you want, and you’ve asked for it, why do we need a third question? 

Question 3: Should I actually believe this?

Always with every AI response of ever, you need to ask one more thing: is this right?

We covered hallucinations properly in H is for Hallucinations, so I won’t repeat all of that here. But it’s worth naming in the context of questions specifically, because the way you ask can affect the answer you get in ways that aren’t always obvious.

The factual check is the obvious one. Numbers, statistics, names, claims that sound authoritative. If you can’t see where it came from, be suspicious. Copilot should point you to sources when it’s drawing on real data. If it isn’t, ask. 

Also – Copilot will lie to you about stuff that it can do. I recently asked it if it could add 10 minute buffers after each of my meetings. It answered me confidently. But… 

When I asked it how it would go about doing this, Copilot fessed up that it couldn’t actually do this thing: 

“Great question — and I’m glad you asked before hitting “yes,” because this one’s a bit nuanced.

👉 I can’t actually implement that behaviour directly via the calendar instruction system.“

Always push back! 

There’s a subtler version of this that I find really interesting: questioning the question itself.

If you ask Copilot “what are the best ways to run a team meeting?” you’ll get a reasonable answer. But you’ll also get an answer shaped by whatever assumptions are baked into the question, and into the model’s training. What counts as “best”? Best for whom?

The same applies to anything that touches on people, culture, or judgement. AI models are trained on human-generated content, which means they can quietly reflect the same biases and blind spots that humans have, just at scale, and with more apparent authority. 

I’ve been thinking about this more since reading The New Age of Sexism by Laura Bates (mentioned in L is for LLM) – it’s a good, uncomfortable read if you want to go deeper on this.

You don’t need to interrogate every Copilot response like a conspiracy theorist. But for anything consequential like a decision, or a document that’s going out, or most importantly; an opinion you’re forming, it’s worth asking: is this actually right, or does it just sound right? 

And is the question I asked already pointing towards a particular answer?

The three questions, in order

Before you open Copilot: What am I actually trying to achieve?

When you write your prompt: Have I given it enough to work with?

When you read the response: Is this right – and was my question a fair one?

None of these are complicated. But the first one especially tends to get skipped in the rush to just get on with it. And that’s usually where things go wrong.

TLDR

Getting the most out of Copilot isn’t just about how you write prompts, it starts way earlier than that. 

First ask yourself what you actually want before you ask Copilot anything. Ask it clearly, with context. And then question what comes back; both for accuracy, and for the assumptions baked into your question in the first place. Three questions. In that order. Every time.

What’s your most expensive “vague prompt” moment? The one that cost you the most time when Copilot went completely off in the wrong direction? Drop it in the comments.