Free Copilot Prompt Generator Agent: Why I Built It

Prompt Generator is a free Microsoft 365 Copilot agent. Describe what you need in one rough line and it writes two or three ready-to-paste prompts using nine proven frameworks, then tells you which to run first. It is available now, and this post explains why I built it and what it will and will not do.

There is a moment in every Copilot enablement session I run. Somebody has just watched a well-built prompt produce a board-ready summary from a pile of meeting transcripts, and they ask the question everyone in the room is thinking: “So what should I actually type?”

That question is the whole problem with Copilot adoption, and it is not a Copilot problem. The licences are bought. The agents are visible in the navigation. The thing missing is the thirty seconds of structure between a vague intention and a request that Copilot can do something useful with.

The standard answer is a prompt library. I have built several for clients. They help for about a fortnight, until someone needs a prompt that is not in the library, and then they are back to typing one vague sentence and concluding that Copilot “does not really work for us”.

So I built something that writes the prompt for you.

What Prompt Generator does

Prompt Generator is a declarative agent that installs into your Microsoft 365 tenant and runs inside Copilot. You open it, type what you are trying to achieve in plain language, and it does three things.

1. It reads your rough request and picks the two or three prompt frameworks that fit the task, using selection rules I developed in live training sessions.
2. It writes a complete prompt for each framework, in its own copy block, with no placeholders. No square brackets, no “insert your detail here”. If a value is unknown it writes the most plausible one and tells you which line to check.
3. It closes with a single line recommending which prompt to run first and why.

You copy the one you want, paste it into a fresh Copilot chat and send it. The prompt is written to pass what I call the paste test: it runs with no edits and no stray formatting characters.

Why frameworks, and why nine of them

I am sceptical of most prompt engineering content. A great deal of it is folklore dressed up as method. But a small number of structures do hold up under repeated use, because each one forces a specific piece of missing information into the request before it is sent.

The one that matters most for Copilot is Microsoft’s own GCSE structure: Goal, Context, Source, Expectation. The Source line is the big lever. “Summarise the project” gets you a generic answer. “Summarise the project using the Migration Steering Group meeting transcripts from September and the status report in the Programme Office SharePoint site” gets you something you can send. Most people never write the Source line because nobody told them it existed.

The other eight earn their place for tasks Copilot handles less well by default: persuasion, strategy, phased plans, content with a particular voice, technical accuracy.

Framework	Best for
GCSE	Everyday Copilot work grounded in named files, emails, meetings or chats. The default.
RTF and APE	Quick transactional outputs: emails, summaries, short rewrites, comparisons.
BAB and CO-STAR	Persuasion and content with a defined voice: pitches, change narratives, LinkedIn posts.
CARE	Reports, retrospectives and stakeholder updates with outcomes and evidence.
CRIT	Strategic or underspecified work. The AI interviews you first, like a discovery call.
RISE	Phased deliverables: implementation plans, migrations, training outlines.
RODES	Training content and technical briefs, with a built-in sense-check step.

Nobody remembers nine frameworks under time pressure. I do not expect them to. That is the point of putting the selection logic inside the agent rather than on a laminated card.

What it looks like in practice

Take the request in the screenshot above: “Help me write a prompt to pitch our new data governance service to a CFO.”

The agent produced two prompts. A BAB prompt, because a pitch is a persuasion task and the before-and-after contrast is what makes a CFO lean in. And a CRIT prompt, because a real service proposition has commercial detail the one-line request does not contain, and CRIT instructs the AI to interview you for it, capped at five questions, before it writes anything.

It recommended CRIT. Its reasoning: for a genuine proposition rather than a draft, the interview step gathers the pricing, the client context and the objections that a good pitch depends on. That is the same advice I would give in the room.

Other requests I have been using it for during testing: summarising a six-month migration for a steering group, planning a Copilot rollout across three business units, and pasting in a prompt that was not working and asking it to make it better. That last one is quietly the most useful. It identifies the nearest framework, rewrites to that standard and offers one alternative.

What it will not do

I would rather you knew the limits before you install it than after.

• It needs a Microsoft 365 Copilot licence. The agent is free. Microsoft’s licensing and any consumption charges are not, and seeing an Agents entry in your navigation does not confirm you have one.
• It needs an installation route. Either your organisation permits custom app upload in Teams, or an administrator publishes it from the Microsoft 365 admin centre. Both routes are covered in the install guide. Neither is guaranteed to take two minutes.
• It writes prompts, not answers. The prompt still runs in Copilot against your data, your permissions and your organisation’s policies. Check any value the agent has assumed before you send it.
• It will not make a bad request good by magic. If the input is too thin to write something specific, it asks one clarifying question. One. If you ignore it, you get a less specific prompt.

On data: the package declares no external actions and no connectors. It runs entirely inside Copilot, so your prompts and responses stay under Microsoft’s service arrangements and your own retention policies. Nothing comes to me. Review the permissions shown at install time, as you would for any agent.

How to get it

Three steps.

1. Download the ZIP from the Prompt Generator page and leave it zipped.
2. In Teams on the web, go to Apps, Manage your apps, Upload an app, Upload a custom app, and select the ZIP. If that option is missing, send the ZIP and the install guide to your Microsoft 365 administrator.
3. Open Prompt Generator from your Copilot agents list and type your first request.

The step-by-step install guide has screenshots for both the personal and the administrator route, plus a troubleshooting section for the usual tenant policy surprises.

If you adapt the agent for your own organisation, and I hope some of you will, keep the attribution line and link back to gethynellis.app. The frameworks and selection rules come from my Copilot enablement work and from AI That Actually Works, and I would like people to be able to find the latest version.

The take

Copilot adoption does not stall because the technology is weak. It stalls at the moment someone sits in front of the chat box and does not know what to type. Prompt libraries treat the symptom. A tool that writes the structure for you, from whatever rough line you give it, treats the cause. That is what Prompt Generator is for. Try it on something real this week and see whether the first answer back is better than the one you usually get.

Gethyn Ellis is a Microsoft Data Platform MVP and Microsoft Certified Trainer, and founder of GRE Solutions Limited. He is the author of AI That Actually Works and delivers Microsoft 365 Copilot and agent enablement for organisations across the UK and Europe.