Transcription
Usually, I do not use such phrasing, but today we will consider that very single golden formula that you will ever need to master prompt creation for chat GPT or other language models. So let's begin. I have over 13 years in entrepreneurship, digital marketing, and blogging. If you are like me, well, maybe not the hairstyle or mustache, but you also know that prompt creation for neural networks is one of the most important skills of the current decade that needs to be mastered. But you are not entirely sure why some prompts generate very general results, while others give you exactly what you are looking for. I have spent hundreds of hours watching videos, guides, courses, and instructions on prompt engineering, applying the acquired knowledge in my practice, and teaching it in many live streams, conferences, and in my speeches. And therefore, in this video, I share six basic blocks that make up any good prompt, so that you can use it to create high-quality results. Firstly, it is critically important to know not only what these six components are, namely task, context, examples, role, tone, and format. Well, and to understand that there is an order of importance in using these elements. And to show you what I mean, let's enter a simple prompt. I am a 34-year-old man weighing 70 kg. Give me a three-month training program. The first part is context, followed by the task. And the reason why the task is higher in the hierarchy than the context is due to simple logic: if we give a task without context, we will still get some sensible answer. But if we give chat GPT context without a task, then nothing good will come of it. In other words, it is mandatory to have a task in your prompt. But at the same time, it is also important to include well-crafted context, examples, role, tone, and format. When you think about writing your prompt, just go through this mental checklist. Using this form will be a constant reminder for you to include enough relevant information to improve your basic prompt. And as you will see in the following parts, not all of these elements are always needed for a good prompt that produces a great result. By the way, so as not to miss all my free live streams, in which I talk, among other things, about prompt engineering methods, earning on neural networks, or developing your own thinking in this very fast era, subscribe to the Telegram channel, there is a lot of useful information there, and there are already over 10,000 of us. You will find the link, of course, in the description and pinned comments. Now let's break down each of these blocks with specific examples and start with the task. Also, along the way, I will add interesting prompt engineering methods for you. Don't miss any of them. The main rule is to use an action verb and start the sentence with it. For example, generate, give, write, analyze, summarize, and, of course, clearly define your goal. This can be a simple task, such as creating a three-month training plan, or it can be something multi-stage, for example, creating a content plan based on a company brief, which first requires analyzing and identifying the target audience, then understanding their needs, and only on this basis creating posts for social networks. If you don't know where to start, then start from the end. By the way, it is precisely the skill of forming a quality, final result that is the fundamental expertise now in prompt creation. This is why neural networks cannot replace experts. Well, at least not yet. The second component is context. And it is the most difficult to compose correctly because the amount of information you can add to it is absolutely infinite. Therefore, let's simplify everything with three questions. First, what does the ideal result look like? Second, what experience do we already have when working with this task? What do we have, what have we tried, what didn't work, and also what situation are we in? So, returning to the training example. We will have me, a 34-year-old man weighing 70 kg, who wants to lose 10 kg of fat in the next 3 months. I only have time to go to the gym twice a week for one hour each time. Give me a three-month training program to follow. Prioritize muscle groups such as chest, shoulders, and deltoids, as well as a training system that maximally preserves muscle mass, because I don't want to be skinny at the end of this three-month race. The global essence here is to limit the infinity of options that any neural network possesses. There is one excellent method that I highly recommend using. It is to ask the neural network in the prompt itself to ask you the questions necessary to complete the task 300%. And here we will need the next point very much, namely role. Essentially, it is who you would like your chat GPT or any other language neural network to become. After all, you would hardly want to get a training plan from a cook. Moreover, you can even use the names of specific people, but this will only work if you do not specify your neighbor, unknown to the neural network, although perhaps an incredible expert in a certain niche, but those about whom there are as many publications, thoughts, and books as possible on the internet, so that the neural network can rely on them. This point will drastically change your result. And to find the role you need, I propose a three-step system. You describe the task, ask the neural network to determine which areas of expertise would best handle this task. Select these areas and ask to find experts in them who are most well-known to neural networks, have a large number of publications, and meet your criteria. And as the third point, simply add their names and perhaps even specific works. Prompt. This works even with fictional characters. So if you ask to write a training plan from Hulk, it will also work. True, it's not a fact that you will be able to understand what he is saying. Let's move on to the examples block. Basically, almost all research on large language models, i.e., LLMs, has shown that using examples significantly affects the final result. Well, let's start with a simple example. Suppose you are filling out a resume, want to change jobs, and statistically this is very likely, and you have some poorly written point in your resume. We can ask the neural network to rewrite it, improve it, using a simple structure. I achieved A with measure B, which led to C. And this is, by the way, a working practice. For example, I reduced the incidence of industrial accidents by 10% by implementing new protocols written by a neural network, which led to saving seventy fingers per year. Or a slightly more complex example in the same resume system. Based on my own resume, answer the question: what is my weakest point? And use the answer structure. This is situation, task, action, and result. And by adding this start, we replace a real example of resume usage, because the neural network will follow a specific methodology. For example, having found some amazing training plan for another person, you can attach it as an example for the neural network and get an output in the same structure. And if you want to create well-written posts that are similar to yours, attach many of your posts that you wrote yourself. And there are many ways to use such examples. But it is important to understand that not every prompt requires adding examples, but when they are added, it will be better. The fifth component is format. And my advice here is simply to close your eyes and visually imagine how you would ideally like to see the picture of the final result. Just don't imagine a million likes under your post, because in that case, it will simply draw you a picture. One of the most underestimated formats is the table format. It works especially well if you use neural networks for information retrieval, comparing pros and cons, as well as simply structured output, so that it is convenient for you to find any of the elements, be it your travel itinerary, daily schedule, or a comparison of which phone is better to buy. Other common formats include emails, bulleted lists, code blocks, but the one that is most useful and works full-time is Markdown. It works especially well if you are summarizing any text, especially raw text, and simply ask it to format it in Markdown. Then, at the output, you will get a very structured output with headings, subheadings, and texts, which, in essence, will simply be an article. Well, and with this task, of course, agents are now best at it, which at the end, in the format of output, provide not only documents but also entire websites. You can watch more about AI agents on my channel under this video. If you don't know which format will be better, ask the neural network to present the answer in several formats. And let it think for itself which of these formats will be the best. For example, a Markdown table or a graph, maybe even an illustration. After all, neural networks have become very good at this. And let's go through the last component - tone. Before we put it all together, the good news is that tone is easy to understand. Use a formal, informal tone of voice, witty response, enthusiasm, sarcasm. Let the neural network sound optimistic or pessimistic. Well, the bad news is that we usually don't remember all these high-flown epithets when we write a prompt, so you can simply convey the feeling you want to get at the output to the neural network. For example, I am writing a message to my client, with whom I am currently working. I want to be taken seriously without appearing too arrogant or causing embarrassment. Give me a list of five keywords for tone that I can include in the prompt for chat GPT. And the neural network itself gives us the desired tone, which we can add to the prompt for generating such text. What is the power of a basic prompt? If you write very simple prompts, such as tell me, how much is 1 + 1, then the answers will be corresponding. But as soon as we add to our simple prompt about 1+1 the role of a storyteller, the task to write an epic story about what 1+1 is, context, as well as output format, we get a completely different result. You can pause and read to feel the difference. And now you know how to create prompts for neural networks at a basic level. Therefore, I hope that you will not miss the next video on fifteen advanced prompt engineering methods by subscribing to this channel. To stay updated on live streams, be sure to add me on Telegram. And I invite everyone who is developing to my club. The world of neural networks does not stand still, so we will see you very soon. Bye-bye.