AI Tools Academy
Microsoft Copilot 0/20

Phase 2 · Microsoft Copilot · Level 1 · Foundations

Prompting Copilot with work context

Walkthrough · 11 minLast checked against the live product: 13 July 2026

30-second recall from earlier lessons
A colleague pastes a signed client contract into their free, personal ChatGPT account to get a quick summary. What's the core problem?
Which of these is safe to type into a free, personal AI chatbot?

By the end, you'll be able to…

  • Write a grounded Copilot prompt that names a source, a task and a shape
  • Point Copilot at a specific file using the '/' trigger in the Microsoft 365 apps
  • Explain why Copilot only ever sees what your own account can already open

Why it matters

The everyday prompting habits from Phase 0 all still apply to Copilot. But Copilot has one idea that's special to it: grounding, meaning you point it at your own real content so it works from that instead of guessing. Getting comfortable with grounding, and with the little '/' trick that names a file, is what turns Copilot from a generic chatbot into something that actually knows about your work.

Everything you already know still holds

Nothing from Phase 0 gets thrown away here. Be specific about what you want. Give context. State the format and the tone. Treat the first reply as a draft and refine it with plain feedback. If you've done Prompting that works, you already have most of what you need: role, task, context, format still carry the load.

So this lesson spends its time on the one thing that's different about Copilot: grounding.

What grounding means, in plain terms

Grounding means pointing Copilot at your own real content (a document, an email thread, a spreadsheet, a meeting) so the answer is built from that, rather than from general knowledge or a plausible guess. The difference is stark:

  • Ungrounded: "Write a project update email." (Copilot invents a generic one.)
  • Grounded: "Using the notes in this document, write a project update email." (Copilot works from your real notes.)

Grounding is what makes Copilot feel like it knows about your work rather than work in general. And here's the limit to keep front of mind: the grounded, works-with-your-own-content behaviour lives in the paid Microsoft 365 Copilot licence inside Word, Excel, Outlook and Teams. The free chat and Copilot Chat can only work from what you type, paste or upload; they're grounded in the web, not your files.

Grounding often happens automatically

Inside a licensed Microsoft 365 app, grounding is frequently automatic. The app you're in is the context:

  • In Word, Copilot already sees the document open in front of you. "Summarise this" means this file.
  • In Outlook, "summarise this thread" means the email chain on screen.
  • In Excel, it works with the sheet you're looking at.

You don't do anything special. Being in the app supplies the grounding.

Referencing a specific file: the "/" trigger

You can also point Copilot at a file you don't have open. In the Microsoft 365 apps and the work chat, typing a forward slash (/) in the Copilot box brings up a search where you name a specific document, email or person, and Copilot pulls from that. Microsoft documents the current steps in Refer to specific files in Microsoft 365 Copilot.

So instead of hoping Copilot guesses which report you mean, you type / and pick it:

Name the file with '/' (paid licence)Copilot
Draft a summary based on /Fernway project brief, focusing on the feedback-process problem. Give me five bullet points and a one-line recommendation.

Why this works: The '/' trigger removes all ambiguity about the source, so Copilot fetches the exact file you pick rather than guessing. Naming the focus and the format on top gives it a target it can hit first time.

This is powerful because it lets Copilot join things up, pulling from a document, an email and a meeting into one answer, each source named with /.

Why Copilot can only see what you can

The licensed Copilot is plugged into your organisation's Microsoft 365 (your files, email, calendar and meetings) and it respects the permissions you already have. Two things follow, and both matter:

  • Copilot can only see what you can already see. It never unlocks hidden files or other people's private folders. If you can't open something, neither can your Copilot. That's a safeguard as much as a limit.
  • The answer is only as good as the source. Vague notes in, vague summary out. Grounding doesn't invent quality; it works from what's there.

The worked example: from ungrounded to grounded

Watch the same task improve as we add grounding and shape. Start with the lazy version:

Before: ungrounded and vagueCopilot
Write an update about the feedback project.

Why this works: With no named source, no audience and no format, Copilot can only produce a generic template, the classic disappointing result. There's nothing here for it to aim at.

Now name the source, the task and the shape:

After: grounded and specificCopilot
Using /Fernway project brief and my email thread with Priya, draft a short update for the operations team on the customer-feedback project: three sentences on where it stands, then the next two actions with owners. Warm, plain, under 120 words. Don't add anything not in those sources.

Why this works: It names the exact source (via '/'), the task, the audience, the length and the tone, so Copilot works from real content toward a describable target. This is the shape every strong Copilot prompt takes.

The second prompt names the three things a strong grounded prompt always names:

  1. The source: which document, email or meeting to work from (/Fernway project brief, "my thread with Priya").
  2. The task: what to do with it (summarise, draft, list the actions).
  3. The shape: format, length and tone (three sentences then a two-row action list, under 120 words, warm).

Then refine, exactly as in Phase 0: "Shorter." "Warmer." "You missed the deadline, add it." Because the earlier context is still on the desk, each nudge keeps what was working.

Refine a grounded draftCopilot
Good, but tighten it to about 80 words, make the tone a little warmer, and add the 31 August deadline the brief mentions.

Why this works: Concrete feedback on the draft you already have is faster than re-prompting from scratch, and Copilot keeps the grounding and the good parts while changing only what you flagged.

Try it now

Common mistakes

  • Expecting grounding in the free chat. Working from your own files and email needs the paid licence. The free chat only knows what you paste in.
  • Assuming Copilot sees everything at your organisation. It's limited to what your account can already open; it won't surface files you can't reach yourself.
  • Vague sources, vague results. "Summarise the stuff about the project" is weak. Name the document or thread, or use / to pick it exactly.
  • Trusting a grounded answer blindly. Grounding cuts guesswork but doesn't eliminate error. Copilot can still misread or overstate what a source says. It's easy to over-trust a reply because it's grounded in your real files, but "grounded" is not "verified". Check anything load-bearing against the source itself.

Keeping current

The / trigger and the grounding idea are durable, but the exact menus, the wording, and what each tier can reach change often. Microsoft's Refer to specific files in Microsoft 365 Copilot and its Copilot help and learning pages are the places to confirm the current steps. Accurate as of 13 July 2026.