Introduction
Nice to have you with us again for another episode of Digital 4 Productivity. And in this episode, I’m sharing my recipe for the best prompts with ChatGPT. ChatGPT is on everyone’s lips, but in my opinion not yet in everyone’s hands. And the main reason for this is that there are one or two misunderstandings. And so, before I tell you the best prompting recipe, perhaps one sentence beforehand.
Most people use ChatGPT like Google and say: Okay, where I googled before, I’ll ask ChatGPT now. Yes, but Google is not the employee you ask, Google is the librarian. In other words, Google’s mission is to catalog the world’s knowledge. And that’s exactly what they do with Google. In other words, you ask a question and you get millions of answers, because there are now millions of sources on certain topics.
In the last episode I talked about the top ten misconceptions and one of the misconceptions was this one about ChatGPT replacing Google. No, if you’re looking for a replacement for Google, it’s Perplexity.
Perplexity is practically a mixture of Google and ChatGPT. You get a structured answer, just like ChatGPT, but you also get four or five main sources and then another 2030 or so additional sources and suggestions on how to go further into a topic. That’s the idea.
What can ChatGPT do?
Yes, what is ChatGPT capable of? For me, ChatGPT is always a new trainee. In other words, a new employee who you can ask questions, who you can assign tasks to, but who you can also ask to challenge you with questions. So you always see ChatGPT as a new employee.
And as with a real employee, you also need to brief ChatGPT properly. What does that mean? With an employee, it also means that you have to give a proper briefing. I always like to tell this nice joke where the journeyman comes back to the foreman and says: “So, foreman, I’ve painted the windows now. Should I paint the frame now too? Of course, we feel the same way with AI, because we always think of something and that doesn’t mean that it’s understood correctly. And just as it’s important to communicate clearly with an employee, it’s the same with AI.
And that’s why I promised you the recipe. Recipe is also my abbreviation, my acronym for the six points that I recommend for an optimal prompt. It starts with the role. In other words, you start by saying: Dear ChatGPT, you are … And if you want to write an article or create a bot on the subject of relativity theory, for example, then it might be a good idea to have Albert Einstein there. In other words, you would say: You are Albert Einstein. The nice thing is that you can use it to assign roles to people who either no longer exist because they are no longer alive, or to whom you have no access. Perhaps not all of you have the opportunity to discuss matters of faith with the Pope, but if you give ChatGPT the role of saying, for example, that you are the sum of all the previous popes of the Catholic Church and you are perhaps Martin Luther as well, then you could have a political and religious discussion with ChatGPT that would certainly have a very interesting quality, to name just one example. Important: On the subject of “role, don’t say who you are, but think about who or what combination of people would ideally be the ideal person to answer this question or ask this question.
In other words, in both directions, if you have a daughter or son, for example, and the son or daughter has an oral exam on the theory of relativity next week, for example, then you can say: So, you are Albert Einstein, who invented this, and now ask these and these questions in preparation for the oral Abitur exam, for example. So that’s the idea of the role and you can combine them. The second is “e, so “r of recipe, “role. The second of “recipe is “e for “result. Be very clear about what you want. So if you tell the apprentice: “Paint the windows and not “Paint the frames, then he will paint the windows and not the frames. And that’s even worse with computers, because computers are stupid. They do exactly what you tell them to do. So to say very clearly, what is the result? And when it comes to results, especially if you want to write an article, it’s a good idea not to say: write an article, but to proceed step by step, as you would with an employee, to say: make a collection of topics, and then you come back and we look at the collection of topics.
Then I might pick out one or two topics to go into in more depth. Then there’s a headline. Then there are another five or six suggestions. I take a look at them. Then there’s an outline. I look at that again. And in the same way, they should ideally also provide step-by-step results, because they also have a so-called context window. In other words, ChatGPT can now memorize up to 100,000 characters in the paid version, which is quite a lot. But you don’t have an unlimited window to the back and you also get significantly better results with humans and also with AI if you proceed iteratively step by step, for example in an article to say: topic, then title, introduction, outline and then the individual points. So role is the first, second is the result. Third point of the recipe: REZ is the target group. In other words, for which target group are you writing an article, for example? It makes a difference whether you are writing an article for a scientific publication or for your daughter’s or son’s kindergarten magazine in order to bring a topic closer to the children. In other words, you have a completely different target group here, or if you are in marketing, for example, you are writing for young people, for example, you are writing for senior citizens, you are writing for a target group that is a beginner in an area you are writing about, rather in the middle range or super professionals.
Write for a target group that wants to have the whole thing in simple language, in complicated language. So define the target group, i.e. R, E, Z.
Then the next E again for place of publication. Where should it appear, staying with this article? Does it make a difference whether the article is to appear in the Frankfurter Allgemeine Zeitung or the Bild newspaper or whether it is to appear in the Tagesspiegel or the Rheinische Post, for example? In other words, here and ideally you should also provide a URL, i.e. a link, if it is a publicly available publication, and say: This is the place of publication, this is what it looks like, this is the style. You can also include “samples and specimens” here. That would be with the next, with recipe, “P”, samples. In other words, you can say that this article should be written like this and this. And then you can either upload texts or provide a link. I keep hearing: Yes, ChatGPT doesn’t write in our style. Yes, if you have a new employee and you say: Write an article about this and this topic, then it makes a difference whether you tell them that on the first day or whether you say: You, have a look at this magazine here, these four or five pages.
We’ve written about it before: It should be in this style. So here it is clear: the more samples you give, the better. The current state of prompting science – there are indeed scientific articles on the subject of prompting – is that examples are good, not too many, not too few, and the order should ideally be random, then you get the best result, so give samples. Then recipe.
The last of the recipes is T for sound. Just set the tone: Should it be a polite tone? Should it be formal? Should it be a formal tone? You can even include a dialect here. So if you are posting a job ad for a Berlin company, for example, you can say: Write it in Berliner Schnauze, or in Cologne you can say: Write it in Gölsch. This works and delivers very, very good results. At least since the 4 Omni version or 4 O, as most people call it, the quality here is really, really good and it’s getting better and better. But the bottleneck is the same as with employees. Usually the manager, i.e. the person who briefs, always notices this when I receive a result back from my assistant that doesn’t meet my expectations, then I first ask myself the question: Have I clearly formulated my expectations? Did I give a clear briefing? And have I put my assistant in a position to support him to the extent that he can do this accordingly? And most of the time, if I’m completely honest, the answer is no.
Conclusion
So the recipe for optimal prompts is Recipe. R for role, E for result, Z for target group, E for place of publication, P for samples and T for tone.
With this in mind, I wish you every success with AI, in the sense of AI is IK, namely your new intelligent colleague.
See you next time. Good luck.
Yours, Thorsten Jekel.
Also available in: Deutsch