Workshop pack · Community meetup

Build the Meeting Bingo Agent

Make your own AI agent in Microsoft 365 Copilot — one that predicts the buzzwords before a meeting, prints you a bingo card, and scores the meeting afterwards.

⏱️ 45 minutes 👥 Pairs or solo 💻 No coding needed 🤖 Microsoft 365 Copilot

1. What even is an agent?

A normal chat assistant forgets the job you gave it. An agent is that assistant with a permanent job description: you tell it who it is, what it is brilliant at, and the rules it must never break — and it stays in character every single time.

🏷️

Name & icon

Who it is, and how people spot it in a list.

📣

Description

One paragraph telling colleagues what it is for.

📜

Instructions

The rulebook. This is where the behaviour actually lives.

🛠️

Knowledge & actions

What it can read, and what it can reach out and do.

2. Before you start

01A Microsoft 365 account with Copilot, signed in on the machine you are using.
02Copilot open in the browser or the Teams app — agent builder lives inside Copilot.
03This page open alongside, so you can copy the spec and the prompts straight across.
04Optional: an agenda document or a few meeting titles you are happy to use as knowledge.

⚠️ Ground rules

The agent pokes fun at phrases, never at people. No names on squares. Nothing confidential — treat this room as public.

3. Our agent — the spec

Hit Copy on each block and paste it straight into the matching field in agent builder.

🏷️ Name

Meeting Bingo Agent

📣 Description

Creates a bingo card before a meeting, predicts the buzzwords most likely to come up based on the meeting title and attendees, and generates a fun scorecard afterwards.

📜 Instructions draft — confirm before the session

You are Meeting Bingo Agent. You turn ordinary meetings into a light-hearted game.

When asked for a bingo card:
- Ask for the meeting title if it has not been given; use the title, the type of meeting and the kinds of roles attending to predict likely phrases.
- Produce a 5x5 grid of buzzwords or stock phrases, with a FREE SPACE in the centre.
- Keep every square humorous but entirely workplace-appropriate: target phrases and habits, never individuals. No names, no in-jokes about a person, nothing anyone would be embarrassed to have on screen.
- Present the grid as a markdown table so it can be pasted anywhere.

When asked what phrases are likely:
- Give a ranked list with a one-line reason for each prediction.

When asked to score a meeting afterwards:
- Ask which squares were called, then give a score out of 25 and a verdict on a scale from "Focused" through "Mildly Jargonised" to "Buzzword Storm".
- Close with one genuinely useful observation about how the meeting could have been tighter.

Always stay playful, brief and inclusive. If a request would embarrass a named person, reshape it into a comment on the phrase instead.

📚 Knowledge

Meeting names and agenda documents. A folder of sample agendas lets the predictions echo the phrasing your organisation actually uses. Optional — the agent works without it.

⚙️ Optional action

Let the agent read meeting details from Microsoft 365, so "bingo for my 3pm" works without you typing the title. Leave it off to demo on invented meetings.

4. Build it, step by step

01Open Copilot agent builder
In Copilot, go to where your agents are listed and choose to create a new one. You get a conversational pane on one side and a configuration form on the other. Either works — this guide uses the form so the whole room sees the same fields.
02Name the agent
Paste the name from the spec and add an icon. When everyone in the room builds the same agent, add your initials so you can find yours later.
03Paste the description
This is the shop window — it is what colleagues read when they pick your agent from a list. It does not drive behaviour on its own.
04Paste the instructions
This is the part that does the work: the grid shape, the tone rules and the scoring scale all live here. Read it through once — you will be editing it at step 7.
05Add knowledge (optional)
Point the agent at agenda documents so its predictions are grounded in real phrasing. Skip it if you have nothing non-confidential to hand.
06Add the Microsoft 365 meeting action (optional)
If your tenant allows it, switch on the capability that lets the agent read meeting details. Every scenario below works with or without it.
07Test in the preview pane
Run the first scenario prompt beside the builder and check three things: is it a 5×5 grid, is the centre a free space, and is every square about a phrase rather than a person? If not, tighten the line that covers it and test again. Two or three loops is normal — and it is the most valuable part of the session.
08Save and share
Save the agent and it appears in your own Copilot agent list. Share it with the person next to you and have them run the same prompt — comparing two grids for the same meeting is the best moment of the workshop.

5. When it doesn't work

The agent doesn't appear

Saving and appearing in the list are not instant. Refresh Copilot and check you are in the account you built it in. If a colleague cannot see one you shared, the share has to complete on your side first.

The output is bland

Vague instructions make vague agents. "Be funny" does nothing; "5×5 grid, free space in the centre, markdown table, phrases not people" does everything. Add the missing constraint as its own line and retest.

You're testing the wrong agent

If the answers ignore your instructions entirely, you are probably talking to plain Copilot or to an older agent. Check the name at the top of the chat before debugging anything else.

It named a person

Strengthen the tone rule and add an explicit refusal line: if a request would embarrass a named person, reshape it into a comment on the phrase. Then re-run the prompt that caused it.

6. Three things to try

Run them in order — each one leans on a different part of the instructions. Every prompt has its own copy button.

Scenario 1 · Everyday The Village Hall Committee

You are on the committee planning the summer fete. Tonight's meeting will circle the tombola for forty minutes. Get a card first.

Create bingo for tonight's summer fete planning meeting. Attendees are the treasurer, the raffle organiser and three volunteers.

Swap out any square that could embarrass someone, and give me five spares I can use if the meeting overruns.

What good looks like

A 5×5 markdown table with a free space in the centre, squares such as "let's take that offline" and "we did it differently last year", and not a single person named.

Try this next: ask for the same card in a version a young helper could play along with.

Scenario 2 · Playful Buzzword Weather Forecast

Forget the card for a moment. Ask the agent to predict the jargon climate for a meeting it has never seen — and to justify every prediction.

What phrases are most likely to appear in a leadership meeting about "unlocking synergies across the portfolio"?

Now rank them by how likely they are in the first ten minutes versus the last ten, and tell me why.

Give me the same forecast for a pub quiz committee, so I can see how much the audience changes the prediction.

What good looks like

A ranked list with a one-line reason each — and two visibly different lists for the two audiences. If both look the same, the instructions are not using the attendees.

Try this next: ask which single phrase it would bet money on, and make it defend the bet.

Scenario 3 · Stretch The Post-Match Analysis

The meeting is over. Score it, then push the agent past the joke into something genuinely useful.

Rate today's architecture review from "Focused" to "Buzzword Storm". We hit 11 squares, including a full row across the top.

Based on that score, what one change to the agenda would have cut the jargon in half?

Write a three-line summary of the verdict I could actually share with the group without anyone taking offence.

What good looks like

A score out of 25, a verdict on the named scale, one real observation about tightening the meeting, and a shareable summary that is funny about the meeting rather than about anyone in it.

Try this next: ask for a season table across four meetings and watch whether it invents scores it was never given — a live lesson in grounding.