AI & Copilot
Prompting Copilot in Microsoft Teams, with examples
People treat Copilot in Microsoft Teams like a search box that somehow knows what they meant. Then they get a vague paragraph, blame the model, and go back to scrolling. Most of the time the prompt was the problem.
I am Priya Shah, a Microsoft Teams automation consultant. This essay is educational field notes from my own client work. I am not a Microsoft partner, and Microsoft does not endorse these examples. If your tenant has Copilot enabled, use these patterns as drafts you adapt to your channels and policies.
Good prompts ask for a shape. Bad prompts ask for vibes. The difference shows up in the first sentence of the answer.
What a useful Teams prompt looks like
A useful prompt names the source, the timeframe, the output format, and the audience. In Teams that usually means a channel, a meeting, or a thread. Without those anchors, Copilot fills gaps with confident fog.
I keep a sticky note near my laptop with four lines: where, when, what shape, who will read it. When I skip a line, the answer gets softer and less checkable.
Where can be “this channel,” “the thread under Maya’s post,” or “Tuesday’s launch sync transcript.” When can be “since March 1” or “in the last two meetings.” Shape can be bullets, a table of decisions, or a short email. Audience can be “for the channel owners” or “for someone who missed the call.”
That last part matters. A summary for an executive should not dump every joke. A summary for an engineer who missed twenty minutes may need the unresolved technical forks.
Also say what to omit. “Do not invent owners. If an owner was not named, write Unassigned.” That single constraint has saved me from fake accountability more times than I want to admit.
Prompt pattern: list decisions and owners
This is the prompt I teach first because meetings and channels drown in discussion and under-document decisions.
Example prompt I actually use:
“From this channel’s messages since last Monday, list decisions that were clearly agreed. For each decision, give a one-line statement, the owner if named, and the message date. If no owner was named, mark Unassigned. Ignore brainstorming that was not closed. Do not add recommendations.”
Why it works: it asks for closed decisions, forces a structure, and bans extra advice. Copilot loves giving advice. You often do not want it yet.
Contrast with a bad version: “Summarize what’s going on and who should do what.” That invites invention. “Should” is a creativity magnet.
When I review the output, I scan for verbs like choose, approve, cancel, postpone, and ship. Soft verbs like discuss, explore, and consider are usually not decisions yet. If the list is full of soft verbs, the channel did not decide. The prompt did its job by revealing emptiness.
I once pasted a decision list into a standup and watched three people say, “Wait, we did not agree to that.” Perfect. The prompt surfaced a false consensus early instead of after a vendor email went out.
Prompt pattern: draft a follow-up under 120 words
After a messy meeting, someone still has to write the follow-up. Copilot can draft. Humans should still send.
Prompt I like:
“Using only this meeting’s transcript and recap, draft a follow-up message under 120 words for the project channel. Include: what we decided, what remains open, and the next checkpoint date if mentioned. Friendly tone, no slang, no praise paragraphs. Do not invent dates. If a date was not said, omit it.”
Word limits matter in Teams. Long follow-ups get skimmed. Short ones get answered.
Bad contrast: “Write a nice follow-up about the meeting.” Nice is undefined. About the meeting is undefined. You will get fluff.
Another bad contrast: “Write a detailed status email covering everything discussed.” Everything discussed includes side quests. Your readers will miss the decisions buried in paragraph four.
I edit every draft for names and tone. Copilot can mishear a nickname. It can also sound oddly formal in a channel that talks like humans. Your job is the last mile.
If your tenant has Copilot enabled but your meeting had no transcript, do not force this prompt against empty context. Say so. “We do not have a transcript; please list what you remember” is a human task, not an AI task.
Prompt pattern: list open questions
Open questions are the unpaid debt of meetings. Teams that track them ship with fewer surprises.
Prompt:
“From this thread, list open questions that were asked but not clearly answered. Group them by topic. For each question, note who asked it if visible, and whether anyone promised an answer later. Exclude questions that were rhetorical or already resolved later in the thread.”
This prompt is ruthless in a good way. It separates curiosity from blockers.
Bad version: “What are people confused about?” Confusion is subjective. You will get thematic essays.
I use open-question lists before I build automation. If the same three questions appear every week, that is a bot or a pinned FAQ candidate. If questions are unique each week, automation will look busy and help nobody.
- Ask for unanswered questions only.
- Ask for grouping so the list is scannable.
- Ask whether someone promised a follow-up, because promises are half-answers.
- Exclude rhetoric so the list stays short.
When the list comes back empty, celebrate. Empty can mean the thread actually closed loops. Or it can mean people stopped asking hard questions. Context tells you which.
Bad prompts, side by side with better ones
Bad: “Catch me up.” Better: “Summarize decisions and open questions in this channel since Thursday for someone returning from leave. Bullets only.”
Bad: “Make this sound executive.” Better: “Rewrite this channel update for a director audience in under 90 words. Keep numbers. Remove names of junior contributors unless they own an action.”
Bad: “Is this a good plan?” Better: “List risks explicitly mentioned in this thread. Do not add new risks. Quote or paraphrase the source message for each risk.”
Bad: “Create tasks for everyone.” Better: “Extract action items that already have a named owner in this meeting. Format as owner, action, due date if stated. Skip items without owners.”
Notice the pattern. Better prompts constrain invention. They ask Copilot to organize evidence, not to become the project manager.
Also avoid prompts that demand certainty the source cannot support. “Confirm legal approval” when legal never posted is a trap. Ask “Was legal approval mentioned? Quote it or say Not found.”
The prompt that failed until I rewrote it
Last winter a logistics client asked Copilot in a busy operations channel: “Who is blocking the carrier switch?”
The answer named three people. Two of them were merely asking clarifying questions. One had been on vacation. The channel lead nearly pinged all three with a stern note.
We paused. Rewrote the prompt together.
New prompt: “In this channel since February 1, find messages that explicitly say the carrier switch is waiting on a person, team, or approval. Quote the relevant sentence. List the named blocker only when the message uses waiting, blocked, pending approval, or cannot proceed. If no such message exists, say No explicit blocker found. Do not infer blockers from questions alone.”
The new answer returned one quote: a finance message saying pricing approval was pending. That matched reality.
The failure was not mystical. The first prompt invited inference. Inference feels helpful until it writes a name on a blame card.
I keep that pair of prompts in a shared note titled “do not ask who is blocking.” It gets laughs. It also prevents a certain kind of Teams drama.
After we fixed the prompt, we still did one human step: the channel lead confirmed with finance in a thread reply before changing any timeline. Copilot found the sentence. Humans owned the escalation.
How I practice with a team for one hour
I do not start with theory. I open a real channel with sanitized permissions, pick three messy threads, and write prompts live.
Round one: decision list. Round two: follow-up under 120 words. Round three: open questions. Then we deliberately write one bad prompt and watch it go wrong. People remember the failure more than the slide.
I ask each person to save two prompts that fit their job. Support leads save open-question prompts. Project managers save decision lists. Executives rarely need their own prompt library if their reports send clean follow-ups.
If your tenant has Copilot enabled for only part of the company, practice in a channel where everyone present can see the same UI. Mixed visibility turns the session into a support ticket.
End with a rule on verification. Any output that assigns blame, money, or dates must be checked against the source message before it leaves the room. That rule is not anti-AI. It is pro-trust.
Prompting in Teams is a craft you learn by rewriting. Start with decisions, short follow-ups, and open questions. Contrast against vague asks. Keep the rewrite of the blocker prompt nearby. Then let the model organize what humans already said, instead of inventing what would be convenient.