Prompt Engineering as an AI Power User by Igor Pogany

Introduction

So, I've been told I need to teach you prompt engineering in 15 minutes, so this should be fun, because prompt engineering is basically the art of communicating with this damn thing. Now, it could be simplified to the fact that it comes down to what you put in here, okay?

But it could also be looked at through the lens of results, goals, because what matters with this app at the end of the day is if you get something useful out of it, okay? That is the ultimate goal that we will be pursuing in the next 15 minutes, getting something that is useful to you. Not to the person next to you, to you personally. That's what defines a good prompt.

1So in a different way, we could also say that it's the art of communication. And as you might imagine, teaching communication in 15 minutes, especially with a new system that might need other inputs than a human, slightly, It's no easy task.

The Essence of Prompt Engineering

That's why I kind of took a workflow that is brand new. We're going to be combining two techniques, one specialty prompt that generates prompts for you, so a prompt that creates prompts.

That's cool. And then we're going to take a second tool that was published by Anthropic.

Actually, it was published two months ago, but it was kind of clunky to use. They had this Google Collab workbook you needed. a little bit of technical knowledge.

You had to press play, API key. It was basic, but they made it even simpler.

So first, we're going to generate prompts for you. And then we're going to take one of those, and we're going to turn it into a super prompt, if you want to call it that. And that's basically going to show you the potential that hides here.

Understanding the Audience

I just want to start with a little show of hands. Who here has used ChatGPT? I really need to know.

Just the basic version, free version, hands up. I suppose everybody at a talk like this. That looks pretty good.

Second question, more important one maybe for this. Who here has used one of the premium plans? ChatGPT+, Cloud, Opus? About half the people.

OK. And of those people, who has built a GPT ever before inside of ChatGPT+. Okay, again, about half of that, so about 25% of the audience has built a GPT.

Roadmap to Building a GPT

So this is essentially going to be kind of a little roadmap to building a GPT. And I don't want to get too deep into that, but that is definitely something to explore.

You just hinted at the 4.0 model that came out this week. I mean, I would only recommend you watch that live stream. I've created content on my YouTube channel around, like, summarizing everything that happened in there, and it's

Hmm, how should I put this except it's really crazy how high the interest of the public is around this topic. It's actually higher than the interest around GPT-4. And the reason for that is because GPT-4 was behind a paid wall.

If you didn't know this new model, this multimodal one that is better than GPT-4, faster, cheaper, All the tools are better and much more. I'm not going to get into the details of that. It's going to roll out to everybody for free.

And it started yesterday, the rollout. So the first three users just go to chatgpt.com, log in and you have something that is more capable than the best model we had seven days ago. Okay. 1And everybody's going to get to access these GPTs.

That's why this little workflow that I'm about to get into. Oh God. And there goes five minutes of my 15 minute crash course. This workflow is going to help you use it, and it's going to help you build these personalized assistants that are coming. Okay? So let's get into it.

Practical Demonstration

Practical demo, here we go.

So if you're a Plus user, you already have this. We're gonna be working in 4.0, and if you're not, then you might have to wait a few days until you get the free version. We are in luck because you do need GPT-4 or 4.0 for this workflow, because this prompt generator that we'll be using It does need a better model.

Now, I'm not here to explain the differences between them to you, but my recommendation would be if you take this topic seriously, which you probably do if you're watching the recording or you're here in person, then you should also be taking the serious models.

OK, so what's going to be step one? I'm just going to show you. And in the end, I'll provide you links to everything. OK, I have QR codes.

Step 1: Generating Custom Prompts

1Let's take a digital marketing manager. We simply copy this, paste it in here and we run it. Okay.

Let's not worry about what's written in here. That's not the point of today. The point of the results, as I said, so

So what's happening right here? The model is generating 30 different use cases. They're titled, and they come with a simple little prompt that gets something done.

You might be familiar with this. I bet you were on Twitter. You subscribed to some newsletters.

You saw some YouTube videos, and they gave you prompts that you noted down somewhere. This is where your journey really begins, because everything that comes down the line, all these buzzwords that you might be hearing, agents, multimodal, I mean, what do we have? Fine-tuned LLMs. Then we also have automations. That's the big one.

A lot of people, especially business owners, are like, how do I get my hands on these automations? That sounds really promising, right? Klarna, customer support agents, replacing 70% of the workforce. This is where it begins, because these are capsules. These are isolated problems that you can solve, right?

We had this question of, how do we teach people on how to interact with agents? Well, we don't. They know how to interact with a person. And then they also know how to interact with one of these automated bots that banks have.

You know, those annoying things where it's like, press one if you speak English and then you sit there for five seconds after you press one. If you can do that, these agents are going to be way better because they're going to have a database of these little problem solution capsules in the background. And what I want to do today is I want to equip you with some of them. And I want to make you understand how to use these for yourself.

Now, we don't have enough time to teach you how to customize this fully for yourself so it can generate the prompts for you. But I think if you're here, chances are, if you look at this last part of it, this is all I'll say about this. This last part, starting here with professional role. If you only customize this thing and then maybe you look at the goal here. There you go. Here's goals. If you customize these two lines to your job, to your work, then this will output 30 custom prompts to what you do. OK? So this is a very quick hack.

Step 2: Improving the Prompt

It's really a shortcut to generate prompts for yourself. But here's the thing.

If you only give it this, I don't know, let's pick a random one. How about improved conversion rates? A classic marketing problem, right?

Let's look at this. By the way, this is freshly generated. I don't even know what it came up with.

Develop a strategy to improve conversion rates on specific page or channel, focusing on optimizing user journey, reducing friction points, and enhancing value propositions. Now look, that's quite shallow and it doesn't even specify what page or channel we have.

Now everything that's in brackets here, you can replace with your own page. So you can say, my website. Now with the browsing tool, you can even give it the link, for example.

These things you become aware of when you use the tool a little bit. That's quite intuitive, I think.

But then the thing is, what makes this really better is making it deeper. So, prompt engineering.

Let's take a step back. The way we put these in is the way you would talk to a brand new employee. So if you hire a new assistant and you really overpay her, you're going to have some expectations.

You're going to be like, this better help me in my business. If your new assistant comes in and you tell her this, and then you just shut up and you let her do her thing, results are not going to be that great.

Assume she knows nothing about you because she doesn't. You didn't give her more info. It would probably help to tell her, what is your page? What's the name of your brand? What kind of target audience do you have? What products do you have? All of these things, right?

But in order to do that, you need a little bit of domain knowledge. Now in the second step, we're gonna take this prompt and we're gonna kind of skip that again. We skipped the prompt generation process, now we're gonna skip the prompt improvement process. Check it out.

Using Anthropic's Console for Prompt Optimization

We're gonna do this for this anthropic, Console. They launched this tool, I believe, pretty much seven days ago. Again, before it was available in a different form, and there's other workflows to improve prompts, but this is by far the most user-friendly one.

You're going to go to console.anthropic.com slash dashboard. Again, in the end, I'll provide you with the links. But basically here, there's a tool in here. where you're going to just say generate prompt.

And what this does is it basically, it's a simplest prompt generator that is actually sophisticated that I know of. So all you need to do, you need to pick one of the prompts that you might have on your computer or one of the ones generated in here. Let's just do this. Copy. Paste. Generate prompt.

Now I did skip one step, which is you need to log into this page and you need to, you get $5 of free credits. Um, and then you kind of connect your credit cards and you can, you know, load a few dollars. Each one of these prompt improvements that is going to take about 15 seconds costs around around 10 to 20 cents. But I would say that's well, well worth it because look, now it's actually going to go deep.

It's going to define the variables that you might not be aware of and converting, um, and optimizing conversion rates. And here's the point. Let's go back to this analogy with the assistant that I always use when I teach prompt engineering. If you give the assistant the first one-liner, the one sentence without the product name, without the website URL, she's not going to do a good job. She can't.

No matter how smart she is, no matter how hard she tries, she would need to do her own research. But in the context of ChatGPT, she can't even do that, right? With this, just think about it, it's basic communication, it's not technical.

If you tell her all of this, so let's have a quick look at this, how much, we have five minutes left, that's perfect. So, you're tasked with developing a strategy to improve conversion rates on a specific page or channel. The main objectives are to optimize the user journey, reduce friction points, and enhance value propositions. It's pretty much the same thing that we have, right?

But it's just the beginning of the journey because now we start fleshing out what else we need to tell that assistant to get results that are actually valuable. So first of all, here's the specific page or channel you will be focusing on. This looks a little technical. That's because this prompt generator is made for developers, but you can absolutely use it for yourself. As you might intuitively realize, all you need to do here is you need to specify the page.

You paste the URL, so GPT can actually access that. Next up, here's an overview of the current strategy being employed. You insert your strategy, and if you don't have it written down, then prompt engineering is communication. That's what I started with, so you're gonna need to communicate a little bit. You're gonna need to tell it. What is our marketing strategy?

If she wants to know this, this is what a really good assistant would ask. What else do we have here? Here's the relevant analytics data for this page or channel. Heck, we're improving conversion rates. I didn't even come up with that when I was kind of talking about what to give the assistants, but she should probably have some number of conversion rates that you already have to improve them. What else is she going to be improving upon, right? Only makes sense.

And this tool does it for you, right? This cost me 10 cents, but it potentially saved me hours. Next up, there's a few extra data points. And then it also always flashes out everything. I'm not going to go into every single one of them because that will take a few minutes. And it's also different with every prompt.

But that's what this does. It basically takes a step back and it asks the question, what else do we need to know to perform this task really well? OK? Let's see. Maybe let's glaze over it. analyze the current strategy, okay, identify friction points, brainstorm ideas for optimizing the user journey, great, develop a comprehensive strategy, and then wrap it up with analysis. And then it gives you everything back, the friction points, the strategy. Okay, amazing.

Organizing and Applying the Optimized Prompts

So then what you would do, you just copy this thing You open up a new chat. You paste this. And now you can fill it out.

Okay? What is also included in this template, I'll give you a link to this. So I provide this with the sign up to my newsletter. I'll give you a direct link here. But basically down here, there's like a prompt organizer. And I use this. You can use your own system. You can use the Excel sheet. You can use your notes. But this is a really great way of kind of organizing this. I'd be like, I would name it, and then I have it in here, right? And then I can access it at any point in time. And if I want to improve a conversion rate, I just copy here, paste in here, and I always have it because, you know, this, let's be real, anybody who uses this a lot, there's no folders, there's no organization, it's a bit of a mess, right? So you kind of want to, one of the main things you want to do right away when you start using these tools is you want an external tool that saves all these prompts, the extensions and stuff. And they change so often, I still prefer having like something like Notion or your Notes app or something like that. 1But basically, now we have a comprehensive message that we can give to our assistant that is getting increasingly incapable. That's what all these updates, all of this space race in the AI industry is about.

They're making the assistants better and better. But what you need to do is you need to learn how to communicate. And it doesn't matter that you might not be as skilled as you want to in real life with communicating with humans. This thing doesn't push back. It doesn't have emotions. It just wants to know the details of what you're trying to achieve.

Conclusion

this two-step workflow of one find the prompt i have my little generator that can help you kind of ideate you can customize it to generate a lot of prompts it's probably the most user-friendly way of getting like a lot of prompts that are custom to you and not just generic so one you find a prompt two you improve it with this prompt optimizer And then you throw it in here and you fill in the variables. And when you run this, you're actually going to get results that are customized to you. You're not going to get this simple responses where people dismiss the technology very quickly because they feel like it's not for them.

If you give it all the analytics, if you give it the data, heck, if you give it a link to your website, Maybe you could even, like, if you learn more about this topic, you're going to realize, hey, it would be really good to give it the target audience. And maybe it also wants some historic data. Maybe you can prepare data that is not just the conversion rate today, but you can give it a whole file of the conversion rates across the past 12 months. All of this is going to give it more information so it can make better decisions.

And then you just run it. So if I run it here, it's not going to get any productive output because I didn't fill in all these variables. But I believe that you could do this yourself. You could give it a link to the page of your website, to the website. And that is really the workflow.

So step one, find prompts that work for you. Step two, improve it with this prompts improver.

Now, how does this link to this topic of GPTs that I alluded to in the beginning? These GPTs.

I'm gonna wrap up here, but these GPTs, they're essentially specialized versions of this general purpose model. And these specialized versions do one thing.

Now, if you care about all the nitty gritty details and the updates, you can go on Twitter, you can go on YouTube, you can find all the updates there, but basically, Today, actually, a screenshot leaked of the new GPT building interface. The main thing they're changing, for anybody who's built GPTs before, is that they're including subsections in the instructions.

I think they call them blocks. And these blocks are always for one task at a time.

So you're going to, let's say, set up an email assistant. And then you're going to have maybe 10 subsections where it's like, one of them is the classic, I want a refund email. Second section would be the classic, I'm lost, where do I find a link to the e-book that I bought, email?

Whatever it might be, depends on your business, right? But there's going to be different use cases. And these blocks are going to be different prompts.

So what does it come down to building a GPT? Well, one of the first steps is going to be, what are the different prompts that you wanted to execute? What are the different tasks that you wanted to execute?

Well, now I showed you that if you identify them, you can use this tool to improve them. You fill in a few blanks. and you're halfway done with building the GPT that can do 10 different things, right?

And the cool thing is, it's not implemented yet, but some other tools have this already. If you start using this, and by the way, a week ago, only the paying users could use this, but within the next few days, all of the 100 million users are gonna have access to this GPT. So this stuff is about to become really relevant.

You could send your grandma a GPT and impress her with that if you're into it or something. Mine is actually surprisingly interested in it and into AI. But that's not the point.

The point is that it all starts with the use cases. All of these downstream stuff, agents, automations, GPTs, All of that starts with this journey. It starts by how do I build something that can solve the problems that I care about? So first you need to identify them. It's the first step. And then you need to flesh it out a little bit. Second step.

And I believe this should conclude our little crash course on problem engineering in 15 minutes. These are shortcuts, but I hope they will help you out in your everyday life. Thank you.

Finished reading?