Introduction
Welcome to another episode of TJ’s Technology Tuesday. First, I need to make a correction from the last episode. I was kindly pointed out, when I introduced the topic of programming, that the provider Claude may sound French but is not French — it comes from San Francisco. That is entirely correct. I apologise for that mistake. I had confused it with Mistral. Mistral are the French, Anthropic are the Americans from San Francisco. Nevertheless, I continue to recommend them for programming tasks, especially since they also take the approach of providing ethical AI.
Today, however, I want to revisit an idea on the topic of prompting. What I keep experiencing is that people enter prompts and are not satisfied with the results. That is why I would like to introduce you to a way of thinking that I always call “Reverse Engineering” or “Reverse Prompting”. The underlying idea is that you put in a result and then ask: “What do I need to prompt in order to get this result?” Let us look at this with two examples. First, just this week I ran a presentation training session at a company where we initially worked out: “What does a good prompt look like?”
Write perfect prompts with the Microsoft Prompt Coach
And what a good prompt looks like can be worked out wonderfully using the Prompt Coach from Microsoft. You will find it in the left-hand panel under “Your Agents”. Even if you do not have the paid version of Copilot, you still have access to it. I am currently using my wife’s licence — that is why it says Nicole Jekel at the bottom — and it is definitely a version that is available within Microsoft 365. There you have all agents, and under “All Agents” you also have the Prompt Coach. If you do not yet have it under “Your Agents”, you will find it among the agents created by Microsoft.
Within the Prompt Coach we simply asked: “Okay, what does an optimal prompt look like?” and first asked the Prompt Coach itself. That already produced quite good results. I then followed up and said: “Okay, supplement the following system.” Those who have known me for a while will recognise this: I use the recipe of Role, Result, Audience, Publication Venue, Samples and Tone.
That is how I remember it. The system then optimised the prompt once more. We continued from there by saying: “Write an article about a product.” It initially wrote a rather poor prompt that was not good. It then asked a follow-up question on that basis. It naturally asked because I had set in the custom instructions that it should ask follow-up questions. It had already remembered this. I then asked what the optimal prompt could look like. Eventually we said: “Very good question. How could I optimise this further?” We then created the optimal prompt and wrote a blog article about that product. That was the first starting point.
The next starting point was that we read through the suggestions the system offered. So again “Optimise” and an SEO-optimised, fully written version. We followed the system’s recommendation, and the longer you work with Copilot, the more it adapts to you and the better the results become.
Do not accept the first version — always ask for more
Then again the question: how could I optimise the article? So do not accept the first version, but ask again. It then delivers corresponding ideas. Once I had all of that, I said: “Based on this chat, create a description for an agent that writes blog articles for small businesses.”
That is precisely the idea: if you have a good prompt and a good chat history where you have developed the content, it is like having a trainee who has done something really well and you say: “Remember this and please do it the same way every time.” The custom settings, custom instructions and saving of content can support this. But now I can say: “Create me a description for an agent.”
Reverse Engineering for perfect agents
It has now created a description for me and it is truly excellent. Honestly, I would not have come up with it myself. It then asked me again: “Do you want me to write that now, or would you like an agent definition?” I had a complete agent definition produced, and after having it revised once more, I built it into an agent. That means here I said: “Create me an agent.” And you can create these agents under “Create new agent” — also, incidentally, in the version included with Microsoft 365. I then entered the whole thing under “Describe”. Since there is a limit of 2,000 characters here, you sometimes need to do it in two steps. It then generates a configuration from it, including things I had worked out there that I certainly would not have formulated as well myself. An important point when it comes to agents: all websites, i.e. the internet, should generally be switched off, and you should specify that the agent should only use the provided sources and nothing else. Less is, in my experience, more. But Code Interpreter — so that it can analyse and work — yes please, images too, and then it produces the corresponding suggestions for me.
Yes, we set it up that way, and when I check, we simply said: “Generate an article from this.” And that was then the result, where we said, with this bot we said: “Write me an article about Aseptoman Pro.” And that is a specific product of that company, and there is a website for it too. And then what comes out wonderfully in one go is a wonderful article with a meta description and also with FAQs, because we had included that as well. AI tools tend to work well with questions and answers. So if you build a bot — an agent — for the topic of blog posts, it is a good idea to say: at the end, always include five Q&As, or ten. And the client was very, very impressed by how well it all worked. So the idea is, especially when building agents, to say: you have great results somewhere, you have great prompts, and then you say: “Come on, do some reverse engineering — build me an instruction for a bot from this, build me a perfect prompt from this.”
I can also say this at the end of a long discussion: “Right, we have now arrived at this result. Please give me a prompt that would take me directly to this result.” That means not saying: “I have a prompt, let’s see what comes out”, but rather “I have a result — and that can also be a real, existing result.” In other words, you can upload an article and say: “Okay, describe how you would prompt this article.” And then you get the “Prompt for it”. I find it always very, very exciting: when you let AI reverse-prompt results, something quite different comes out than what we as humans would have thought to prompt. This works wonderfully for images too. And for images we naturally look at my current favourite model, namely Google “Nano-Banana”. You will find this within Google Gemini. What most people do is open a new chat and head straight for the banana — generating an image. Do not do that. First, get the banana out. It is important to have Thinking activated with three Pros.
How to create perfect prompts for image generation
And now here is what you do: you say: “Describe this image as the basis for a Nano-Banana-Pro prompt.” So I uploaded a photo of my wife from an event and said: “Describe this to me.” And now a detailed image analysis follows — a woman, her pose, her clothing, the background, central elements, ambient details, lighting style and so on. And then comes a prompt suggestion for image generators. That means it shows me an image analysis, gives me a prompt suggestion — suggestion one in English, because it knows that models work better in English — gives me a negative prompt, and even provides notes on what I should watch out for: for example, that I should look into the text topic, that the face is relatively small in the photo. So there could be difficulties there. And then I said: “Create a similar image with the same woman from the uploaded photo in a red dress, 16:9 landscape format, with the prompt ‘Mission 1’.” When I compare: that was the original photo of my wife, this is the dress in red, and it looks pretty good. So the idea here is: always upload a reference image that is as similar as possible.
So even if you have an image idea — perhaps your shop with two customers, for instance — but you say: “I don’t want to appear in it myself”, then take a photo of two employees and replace the two with a different person, for example. You can even combine up to 14 objects in these Nano-Banana jobs. Then I said: “Okay, place this one next to her in a similar pose.” That is really just a torso shot from a train — a selfie. No arms, no legs visible. Quite minimal. And I said: “Put him next to her.” And he is 10 centimetres taller. So I give the AI this information. That is also something worth noting: how would the system know, from a photo like that, whether the man is 2.10 metres tall or 1.50 metres tall? It does not. So you should pass that information along. He is 10 centimetres taller, and then it placed him next to her. Those are not trousers from my wardrobe. I do have a suit, but a different one — though that information was not provided, and it filled it in well.
Since the last photo was in portrait format, it placed the result in portrait format too. But that can be changed immediately by saying: “Please use 16:9 landscape format.” Now I want the table it added removed, and the pole. I said: “Okay, remove the table.” Done. The pole might need a second pass. So you get an idea of the logic behind it. The key point is that for images, it is almost even more important to first upload a photo that is close to what you have in mind — one that says: “This is how I envision it and this is what it should look like.” It should be relatively close. You can also say for the image prompt: “How can I optimise this?” You can also say: “Try to replicate the camera settings and the camera from this image approximately.” Or if you know it was taken with a specific iPhone, say: “Okay, was it taken with this?” In other words, have the prompt generated first and then apply that prompt. And this applies both to results you have already generated with AI and results you have generated with DALL-E.
Copilot delivers significantly more useful results compared to ChatGPT
You effectively reverse the process. The same applies if you are in the middle of a prompt conversation — simply open another prompt session and say: “Here I have a document. How would a good prompt look for this?” And then you can copy it back across, perhaps because the chat is getting a bit long, or because you do not want to “contaminate” it with side actions. So again: Reverse Engineering makes a great deal of sense in the field of artificial intelligence too. I hope this was once again a valuable impulse, because as you know, my motto is always “Switch on the brain first, then the technology, and then simply use the technology.” And I often hear: “Well, ChatGPT is much better than Copilot.” But if I brief Copilot properly, and if I access company data in the paid version, then the results in a business context are significantly more useful than what I get from ChatGPT. And the data is also stored with Microsoft in Germany in your own tenant. The AI processing currently still runs in Ireland, but we are also in the European Union there, and the data is in Frankfurt.
Companies want and must avoid shadow AI
So if someone says: “We have enabled Outlook and everything like that, but we have not enabled Copilot, even in the included version” — then I say they would also have to be consistent and disconnect SharePoint and Teams, because the data storage location is the same. I recently had a discussion on Friday with a client who also said: “Actually, you are right — if we disconnect it, what will employees do?” There are studies showing that 80% of employees now use AI, but only 40% of companies provide AI tools. This is called “Shadow IT” and “Shadow AI”, and you definitely do not want that. So if you have Microsoft 365, by all means enable Microsoft Copilot, and then raise awareness and show how to prompt it so well that nobody has any desire to go back to ChatGPT. Perhaps one final impulse: if you create agents that also connect to SharePoint libraries or Teams, for example, you need the paid licence for that. The good news is: only the person who creates them needs it. That means you can give these paid licences to a small team, and they can then distribute these agents via Teams, SharePoint or links.
Conclusion
And employees do not need those paid Copilot licences, which cost another 30 euros per month. For me, however, this Copilot — when used properly — saves significantly more than 30 euros worth of working time. Against that backdrop, it is an investment worth thinking about seriously. In that spirit: brain on first in the decision, then technology.
Your Thorsten Jekel.
Key Takeaways
- Reverse Prompting means presenting a desired result to the AI and asking what prompt would be needed to produce it — rather than the other way round.
- The Microsoft Copilot Prompt Coach helps systematically improve prompts and is included in the Microsoft 365 licence at no additional cost.
- A proven recipe for good prompts is: Role, Result, Audience, Publication Venue, Samples and Tone.
- A successful chat history can be used to directly derive an agent definition — the AI formulates the agent configuration based on the developed content.
- For image generation, it is recommended to upload a similar reference image and let the AI derive the appropriate prompt from it (e.g. with Google Imagen 3).
- Important image details such as height or format specifications (e.g. 16:9 landscape) should be explicitly stated in the prompt, as the AI does not know these on its own.
- In a business context, a well-briefed Microsoft Copilot with access to company data delivers significantly better results than public tools such as ChatGPT.
- Around 80% of employees already use AI, but only 40% of companies provide official AI tools — Shadow AI is the result and should be avoided.
- For creating agents connected to SharePoint or Teams, only the creator needs a paid Copilot licence; other employees can use the agents without an additional licence.
- The conclusion: switch on the brain first, then the technology — Reverse Prompting helps to fully exploit the potential of AI tools.
Frequently Asked Questions
What is Reverse Prompting and how does it work?
Reverse Prompting reverses the usual approach: instead of writing a prompt and seeing what comes out, you present the AI with a desired result and ask what prompt would be needed to achieve exactly that result. This yields more precise prompts that you can then reuse directly.
What is the Microsoft Copilot Prompt Coach and where can I find it?
The Prompt Coach is an agent within Microsoft Copilot that helps optimise and refine prompts. It is found in the left-hand panel under “Your Agents” and is available even without a paid Copilot licence as part of Microsoft 365.
What recipe does Thorsten Jekel recommend for a good prompt?
Thorsten Jekel recommends the recipe “Role, Result, Audience, Publication Venue, Samples and Tone”. These six elements help formulate a structured and effective prompt that gives the AI all the information it needs.
How can you create an agent definition from a good chat history?
At the end of a successful chat, you can ask the AI to create a description or complete configuration for an agent based on the entire conversation. This agent definition can then be inserted directly into Microsoft Copilot under “Create New Agent”.
Why should you upload a reference image for image generation?
An uploaded reference image allows the AI to generate a detailed and fitting prompt that takes into account pose, clothing, background, lighting style and other details. This automatically generated prompt can then be used for the actual image generation, delivering significantly better results.
Which AI model is recommended for image generation?
Google Imagen 3 (referred to in the video as “Nano-Banana”) is recommended, available within Google Gemini. For optimal results, ensure the Thinking option is activated with three Pros, and write image prompts preferably in English, as models work better with it.
Why does Microsoft Copilot often deliver better results than ChatGPT in a business context?
In the paid version, Microsoft Copilot can access company data directly in SharePoint and Teams and works within your own Microsoft tenant. This means responses are tailored to the specific business context, leading to significantly more relevant results than a general tool like ChatGPT.
What is Shadow AI and why is it problematic?
Shadow AI refers to employees using AI tools that have not been officially provided or approved by the company. Studies show that 80% of employees already use AI, but only 40% of companies officially provide corresponding tools — a security and compliance risk that companies should avoid.
Who needs a paid Copilot licence to use agents?
Only the person who creates an agent needs a paid Copilot licence (approx. 30 euros per month). Employees who use the finished agent via Teams, SharePoint or a link do not need their own paid licence — saving companies considerable costs.
How can you use Reverse Prompting for existing documents?
You can upload an existing document or finished article to the AI and ask it to describe the prompt with which that result could have been produced. This reverse-engineered prompt can then be used directly for similar tasks or as the basis for an agent.
Tools & Resources Mentioned
- Artificial Intelligence (AI) — Overview page on AI topics at digital4productivity.de
- Microsoft 365 / Microsoft Copilot — Copilot is included in Microsoft 365; the Prompt Coach is available as an agent at no extra charge
- Google Gemini with Imagen 3 — AI model for image generation, referred to in the video as “Nano-Banana”; prompts in English recommended
- Anthropic / Claude — Mentioned in the video as a provider from San Francisco, recommended for programming tasks and ethical AI
- Mistral — French AI provider, mentioned in the video as a correction reference
