The starting point is different from Claude's
If you've used Claude, you're used to a chat starting from nothing. It knows only what you type, attach, or explicitly connect. Copilot works the other way round: from the moment you open it, it's already grounded in Microsoft Graph, your Outlook mail, your calendar, your Teams chats, and any SharePoint or OneDrive files you have access to. Ask "what's on my plate this week" in Copilot and it can genuinely answer, pulling from your actual calendar and inbox, no attaching required.
Why that's a head start, and a risk
The upside is real: you often don't need to explain background Copilot can already see. The catch is that "can already see" means everything you technically have access to, including old, loosely shared files and folders nobody remembers granting access to. A broad question can surface content that was shared months or years ago and quietly forgotten, not because Copilot did anything wrong, but because it removed the friction that used to keep those old permissions harmless. This is the same point Rollo's AI Act material makes about Copilot's risk being "archaeological" rather than new: the exposure was already there, Copilot just makes it easy to find.
Being deliberate about context anyway
Even though Copilot starts pre-grounded, you can, and often should, point it at something specific rather than relying on a broad, tenant-wide search.
- Reference a specific file, email, or meeting by name in your prompt, or attach it directly, the way you would in Claude.
- Inside a document, Copilot grounds itself to that document by default: narrower and more predictable than a standalone chat question.
- Use "Add content" in Copilot Chat or a Page to explicitly bring in a file or link rather than hoping Copilot finds the right one on its own. A narrowly scoped question is both more accurate (less noise to sift through) and safer (less chance of surfacing something that shouldn't have come up).
Three places to point that deliberateness
Once you're being deliberate, it helps to know Copilot has scoped levels of its own, the same idea as Claude's profile instructions, project instructions, and per-message context, just under different names.
- This message. Attach a file, reference something by name, or use "Add content." Applies to the one chat you're in, right now. The narrow scoping from above.
- This Notebook. A Copilot Notebook (the closest thing Copilot has to a Claude project) can carry its own custom instructions, set from the notebook's menu, that apply to every chat inside it and stay with the shared material the notebook holds. Share the notebook and colleagues can see and co-edit those instructions too.
- Everything. In Settings, under Personalization, "Customize how Copilot responds to you" applies your preferences to every future conversation: your role, the format you like answers in, terms you'd rather it didn't have to be told twice. Copilot also keeps its own Memory of details from past conversations, separate from this setting but working alongside it. Widen the scope the same way you would in Claude: when you notice yourself repeating the same instruction chat after chat, that's the signal to move it up a level.
Write a short brief anyway
Being pre-grounded is not the same as being briefed. Copilot can see your calendar and inbox, but it still doesn't know your role, your priorities this quarter, or how you like an answer formatted, unless you tell it. The same short brief from Claude's own context lesson works here: a few lines on who you are, what you're usually trying to get done, who reads your work, and the tone and format you like. Drop it into "Customize how Copilot responds to you" if it should apply everywhere, or into a Notebook's custom instructions if it's specific to one piece of ongoing work. Either way, a few minutes spent writing it down beats letting Copilot guess from whatever it happens to find in your Graph data.
Here's what that looks like in practice, short enough to write in a couple of minutes:
I'm an account manager for our top-tier clients. I mostly draft client-facing emails, prep for renewal calls, and summarise account history before a meeting. My manager and the client-facing team read my work, so keep the tone professional but not stiff. When I ask for a summary, default to bullet points with any deadline or dollar figure called out clearly, not buried in a paragraph.
That's four sentences, and it does real work: it tells Copilot your role, your recurring tasks, your audience, and a formatting preference, all things it otherwise has to guess at from whatever happens to sit in your Graph data that day. Drop something like it into Settings once, and every future conversation starts from that baseline instead of zero. Update it when your role changes, rather than leaving Copilot working from an outdated picture of what you do.
The Article 4 connection
Knowing what your AI can see before you rely on it isn't just good practice. It's the practical version of the EU AI Act's Article 4 AI-literacy obligation, which already applies to every organisation using tools like Copilot. For Copilot specifically, that means treating a periodic look at what's actually shared across your tenant as part of using the tool responsibly, not a one-off IT project.
Where this goes next
Good context and good prompting are two halves of the same skill, whichever tool you're using. Context tells Copilot who it's helping and what it can already assume; the prompt tells it what to do with that. The difference with Copilot is that a lot of the context arrives whether you ask for it or not, which makes the deliberate half of the skill, choosing what to narrow, what to brief, and what to double-check, more important, not less.
Try it
Ask Copilot a broad question ("what have I been working on this week?") and note where the information came from. Then ask a narrowly scoped version, naming a specific file, email thread, or meeting. Compare the accuracy and relevance of the two answers. If you have a Notebook, add one line of custom instructions to it and ask the same question again from inside it.