In lesson 2.5 you met grounding: the fact that Copilot starts pre-loaded with your mail, files, and chats, instead of a blank page. We left the mechanics for later. Later is now.
This lesson fills in what actually happens between you hitting enter and Copilot answering, and what to do when a long chat stops being sharp. Once you can picture it, a confusing answer stops feeling random and starts feeling like something you can fix.
There's no number to watch
Ask how big Claude's context window is and you get a clean answer: a published figure in tokens, right there in Anthropic's own docs. Ask the same question about Copilot and there isn't one. Microsoft doesn't publish a context-window size for Microsoft 365 Copilot the way Anthropic does for Claude. There's no meter, no token count, nothing to watch tick upward as a chat runs long.
That's not an oversight, it's a different design. Claude's window is something you load yourself: you upload what you want in, so a visible limit tells you how much room is left. Copilot's context is built for you, automatically, from whatever it finds. There's less to manage by hand, but also less visibility into what's actually in the mix at any moment.
How a prompt actually gets answered
Here's the real sequence, not a simplified version of it. You type a prompt into Copilot Chat, or the sidebar in Word, Excel, Outlook, or Teams. Before that prompt ever reaches the underlying model, Copilot runs it through grounding: it checks Microsoft Graph, which is Microsoft's umbrella term for your mail, calendar, files, and chats, for anything relevant to what you asked. Finding "relevant" is the job of the semantic index, a search layer that understands meaning, not just matching words. It can tell that a question about "the client who was unhappy" might match an email that says "frustrated" or "disappointed," not just the word "unhappy" itself.
Whatever the semantic index turns up gets appended to your prompt. Only then does the combined package, your words plus whatever Copilot found, go to the model that writes the answer. The model never sees your raw mailbox or your SharePoint site directly; it sees exactly what grounding handed it, nothing more.
One rule holds through all of this: Copilot only surfaces what you personally have permission to see. If a file is outside your access, grounding can't pull it in, full stop. That's worth remembering if an answer ever feels thinner than you expected: it might be working exactly right, on a narrower set of files than you assumed.
The strength, and the catch
The strength is real. You don't have to go hunting for a document, copy its contents, and paste them in before asking a question. Copilot already has a route to your actual work. Ask about "the Devlin contract" and, if you've got a file by that name you can access, there's a decent chance Copilot finds it without you attaching anything.
The catch is just as real. Because grounding runs fresh on every prompt, pulling from whatever looks relevant right now, the exact same question can come back with a different answer depending on what was recently active, what got renamed, or which of three similarly named files happened to rank highest that day. A Claude prompt without enough detail is predictably vague. A Copilot prompt without enough detail can be unpredictable, because it's not just guessing, it's also choosing what to ground on. That's exactly why lesson 3.4 told you to name your scope: a specific file, mailbox, or time range turns a guess into a lookup.
Why a long chat still goes sideways
No published token limit doesn't mean no limit. Push a Copilot chat on, message after message, and two things tend to happen. First, the chat itself can hit a hard stop: Copilot will tell you the conversation has reached its limit and ask you to start a new one. Second, and sooner than that wall, the quality can slip. Early instructions get buried under everything said since. Grounding, re-run on each new prompt, can start surfacing different material than it did at the start, because the thread's drifted onto a different topic than the one it was grounded for originally. The result feels a lot like the "drift" Claude users hit with a full window, even though the mechanism underneath isn't the same.
The practical sign is the same either way: the chat starts repeating itself, forgetting something you said early on, or giving you answers that feel a half-step off from what you actually need.
When to start fresh, and how to carry work over
The fix doesn't need a number either. It's a habit: one topic per chat, and a clean start the moment a conversation drifts or the topic genuinely changes. A new chat re-runs grounding from scratch, against your current prompt, with none of the earlier thread's baggage attached.
Starting over doesn't mean losing the thread. Ask Copilot to hand its own work off before you close the chat:
Before we stop, write a short summary I can paste into a new chat. Cover what we decided, where the draft stands now, and anything still open.
Open a fresh chat, paste that in, and carry on. You keep what mattered and drop everything that didn't. If it's work you'll come back to more than once, that summary deserves a more permanent home than a chat you'll eventually close for good. That's what Copilot Notebooks are for, a later lesson in this Part.
Keeping grounding working for you
A handful of habits keep Copilot grounding on the right material instead of guessing at it.
Name your scope whenever it matters. "Summarise our Q2 numbers" can mean five different files; "summarise the Q2 numbers in the Finance Review channel, posted this month" can only mean one. The narrower the ask, the less grounding has to gamble.
Feed it lean files when you are attaching something directly rather than relying on grounding to find it. A plain-text or Word document carries the same information as a heavily formatted one for a fraction of the size, which matters once a file is large.
Match the mode to the task. Quick response is fine for a routine question; save Think Deeper, from lesson 2.3, for anything where grounding has to weigh several possible sources against each other before answering.
And when an answer looks a little thin, ask whether grounding found what you expected it to. If you're not sure, naming the file or mailbox explicitly is faster than guessing at a second phrasing of the same vague question.
Where this goes next
You now know the piece that makes Copilot both more convenient and less predictable than Claude: it builds context for you, automatically, from whatever you're allowed to see, rather than waiting for you to load it in yourself. That's the quiet mechanic behind nearly everything else in this Part.
Next up: Agent Builder, which lets you package a prompt and a slice of grounding into something you can reuse without re-explaining it every time. After that, Copilot Notebooks gives the recurring work a permanent home instead of a chat you'll eventually lose.
Try it
Pick a real question you'd normally ask Copilot about your own work, the kind where the answer depends on a specific file, email, or recent conversation. Ask it two ways: once in your normal, loose phrasing, and once naming the exact file, mailbox, or channel you mean.
Compare the two answers. If the second one came back tighter, more confident, or simply correct where the first one guessed, that gap is grounding working with better information instead of less of it.