Practical Prompt Engineering: Write Prompts That Actually Work

1. Intro: And welcome to my AI prompting class. So in this class, I'll be sharing with you the very best tips on prompting AI that I've learned from over hundreds of hours of experimenting, trying, doing courses from some of the main providers, all of that distilled into this class here, everything with the goal of getting the best out of the latest larger language models we have available right now. So I've gone through a lot of stuff, so you don't have to from full length courses from providers like Google, Open AI, Anthropic. So I've gone through a lot of YouTube videos, podcast, articles, and I've basically distilled all of the best bits into this class, so you don't have to go through the whole process that I went through, in terms of what works, in terms of what works, practically and what doesn't my name is Hans, a content creator, property investor, and former engineer. And like a lot of people, I use larger language every single day, in all walks of my life. And it's no exaggeration to say that this is a revolution that we're seeing right in front of our eyes. So most of us use large language models every single day, whether it's Gemini or JAPT or Claude, but very few of us stop and think at how significant an invention this actually is and how much is changing our lives. When you think of it in the context of history. So just to put it all into perspective, if we look at some of the most major players historically in terms of the techne and how long it took them to reach 1 million active users, we see that Netflix took 3.5 months to reach 1 million users. Is an impressive feat in and of itself. Facebook, when they launched took ten months. You know, getting 1 million active users onto any platform is a really big feat. And what about ChatGPT on its launch? It did it in just five days. And it's even crazier, if you look at the time frames for companies to reach 100 million active users, which only a very few elites select portion of tech companies ever make to this number of active users. And ChatGPT got to this milestone in just 2.5 months. It's absolutely ridiculous this how big of a change this is to everyone's lives. And going forward, it will likely to continue to be. So Dems Habies the founder of Google Deep Mind AI, says that this is one of the most ferocious battles that ever existed in the history of Tech. And I completely agree with him. This technology has such far wide reaching implications, and it's changing pretty much every industry. Nothing's being unaffected by this. And even though almost everyone uses AI nowadays on basically a daily basis, but very few actually stop and think what's the optimal way to get the best out of these large language model, so much untapped potential and people aren't using them correctly or using the best practices. And I'm sure that the skill of really understanding larger language models and the ability to get the best out of them compared to everyone else will be a really key skill in the next five years going forward. And it'll really put you ahead because after all, if you look at the user interface of the average LLM AI model, it basically looks like a search but underneath the hood, it's not even in the same ballpark. They're very, very different things, but most people do still use it like a glorified search engine. By the end of this class, I'm sure you'll level up your prompting game, whether you've not looked into anything at all regarding prompting or you've looked at some articles and tips, I'm sure you'll all be able to walk away and pick up something that you can practically use in your everyday life. And not only the latest tips and tricks, but I'll give you a framework of thinking, a philosophy of how to be a good prompt engineering at your core, because a lot of the tips and tricks that I learned two or three years ago is already out of date, such as the pace of AI. But being able to think from first principles, how to apply these techniques will have applications going forward. And the interesting thing about this class is even though it is specifically about AI LLMs, a lot of the skills you'll learn here will have far reaching applications across all walks of life because what we teach here isn't just specifically how to prompt AI. It's how to break down a problem logic how to communicate very clearly. So I'm sure you'll definitely pick something up, and I'm very excited to share some of my best tips with you. I see you on the other side. 2. Welcome + Class Project: Welcome on board. I'm really excited to have you on this class here. So we'll be going through a lot in this class. And the way that I've designed it is that everything is meant to be super practical because I believe that's the best way to learn, not by, you know, sitting here and just taking things in passively. But taking it away, trying everything yourself, to help you really internalize it and experiment and see what gels with you personally. So just to give you a quick overview of this class, we'll be going through from the very beginning in terms of how LLMs work on high level, under the hood. So it gives you a bit of background perspective and also what some of the limitations might be. And once we've had some grounding, we'll go through the foundations of what makes a good prompt. So we'll have a foundational structure that basically you'll be following with every single prompt. So that'll be our bread and butter foundation, and then something you can build upon with the more advanced techniques. We will look at how you can best get the AI to tailor responses to you personally. Some of the biggest misconceptions and mistakes that people make when using LLMs. Then we'll look at some of the more advanced topics. So this is based on the latest advice from AI engineers that are working at these big model companies. Topics like tailored instructions, few shop prompting, using the LLM to improve its own responses, and how to iterate your own prompts. So that's a lot of info. I'm going to be throwing it at you. And as you go through it, there's no better way to internalize it by having a practical class project that's something you can use your newfound knowledge on immediately and not letting it just kind of sit in your head, you think you understand it, and then you moved on and then things fade away. That's the philosophy we're taking to learn here. And also, if you're continuously trying things as you go along, you're far less likely to be overwhelmed with the load of information helps you digest it. So the class project is very close to my heart as a content creator. So the project is to build your own content creator sea pilot prompt. In terms of platform, it doesn't really matter. So it could be a for a YouTube channel, it could be for a newsletter, Instagram, not too concerned with the exact platform. What I want you to be able to do is by the end of this course, I want you to use AI to help define your niche and also to write five pieces of content that is highly tailored to you. So if it's YouTube, it'll be five usable video scripts in your specific niche, that's highly tailored. So a little preview ahead, you'll be learning how to construct prompts using the framework of context, task, and tone, which every single prompt going forward, will follow this structure, with the AI in such a way where you can strip away all of the assumptions, which inadvertently also reduces the level of hallucinations to get exactly what you want. And you can give you a highly specific info about the task at hand. For example, in the class project, be who your audience is, what your niche is. So when it comes to generating the content, it won't just be a very generic script that anyone could have written. So one great way I use AI as a content creator, in addition to asking it to do tasks is something to bounce ideas off of. So an interesting topic that I hear a lot of people say with AI is, you know, when is it going to replace humans? What can it do that humans can't do? And I feel like this conversation a lot of the time is missing the point. So it reminds me of this Steve Jobs interview that I've heard years and years ago, where he's comparing the energy efficiency of locomotion, basically, the energy efficiency of lots of different animals in the animal's kingdom to a kilometer per unit of body weight. So it's normalized by unit of body weight, and there's lots of stuff on there. There's like horses, there's mice, there's humans, lots of stuff. And if you look at where humans are just from the energy efficiency of walking a kilometer, humans do fairly well, but not the best. There are other animals that are more efficient at moving than us. Like, for example, when you correct for body weight, horses or salmon are far more efficient at moving than us. However, if you take that human and put him or her on a bicycle, that level of efficiency just skyrockets and nothing even comes close to it in the entire animal kingdom. And I feel like AI is a bit like that, but, you know, jacked up to an extreme. So instead of a bicycle, it's like a rocket. Where it's going to enable you to do things you never dreamed possible without it. But it opens up so many possibilities. But whether it's a bicycle or a rocket, it still requires a human to drive it. It still requires that human intuition to pilot it. Otherwise, it will just take you somewhere where you don't know where it's going. So the first thing I want you to do is just to give the project a try using the techniques you already know or how you use AI right now. So ask you to help define your niche, ask it to write you five pieces of content once you've done that, and we just save that in a text editor, Google Docs doesn't matter. But as long as you've got it saved, and the point of the class is we'll go through it. We'll keep trying these different techniques, and by the end of it, I want to compare your outputs at the end to when you had at the beginning to see what the differences are for yourself. Right, now let's really get 3. How LLMs Work and Their Limitations: So a lot of people call tools like Gemini, hatiPT Claude, just AI, which is technically true, but there are very specific application of AI called large language models or LLMs for short. It's a really important dstinction to make because it helps us understand a little bit more about what goes on under the hood and it helps us understand what it's strong at and what are some of its limitations hopefully where we can overcome. So under the hood, imagine an AI assistant is something where the answer has been artificially cut off, and its job is to assign probabilities to predict what the next most likely word is or most likely token. The idea is you would feed it a question, and its job is to find the most likely plausible answer, given that question, what it does is once you've given it a question have a matrix of all sorts of different various words, and it will assign probabilities to each of those words, and it'll build up your answer bit by bit from those probability matrixes. Now, crucially the thing what it doesn't do is it doesn't go into some database, and it brings out a predefined set of answers for certain questions because that would not make it sound very natural at all, and it wouldn't be tailored to you. So that's how you get different answers, even with very small changes in prompts that you give it. So even though let's say you had access to that matrix, and you could see all of the probabilities, you still couldn't, with 100% certainty, predict what it's going to output. And the way it generates these matrices and assigns these probabilities is by training on huge amounts of data. So, for example, if you look at Gemini 3.1, which at the time of filming is the latest Gemini model, for a human to read through all of the training data that it's been through, if a human was reading 247, it would take over 10,000 years to read all of that training data. So it's an insane amount to train up a modern LLM. But these really advanced LLMs, as we know them, have only been around fairly recently. Underlying technology behind them is actually very old. The really early rule based large language models were actually around in the 1960s. And by the 1990s, it moved to a statistical model of text prediction, which is a little bit more similar to what we've got now. But still know exactly what we in the modern day, think of as large language models. You know, think text prediction models in your textap or your Gmail, stuff like this. And, you know, when you type a few words and it tries to complete your sentence, those sorts of things. Real breakthrough came in 2017 when Google invented the transformer, where words can be taken in parallel, which gave it the ability to understand context and give these amazingly tailored responses. And it gives you this perception that it almost understands what you're thinking. So what does all this background stuff mean for us practically? Well, now that we have a very rough overview of what's going on in the background, we can understand a few things in terms of its limitations in the fact that, you know, while it sounds like and it imitates very well in terms of understanding what you think, it doesn't actually understand what you think. All it's doing is taking your input and predicting what could be a plausible answer. All it is is a predictive engine. So I use the analogy of the bicycle or rocket of the mind, and it kind of applies here. Basically, it doesn't care or it doesn't know where you're going. You know, it has the ability to go really, really fast. But it has no awareness of where it's going or why. That's why it needs a human driver to direct it. And also, one of the biggest issues with these AI LLMs, now that we know it basically its primary goal is to give you the most plausible answer, given the context that you've provided it. It means that, on the other end, when you don't give it enough info, it will still try to generate and because all it's doing is trying to do the best job possible with what it's got to provide a plausible answer. Based on its probability matrix. Let's just give an example of an LLM versus a human assistant for a fairly simple task. So here is the prompt. Write an email to my boss asking for a deadline extension on the project. And if you look at the answer, it's pretty polite and it's plausible looking. But if you actually read through it, there's a huge amount of stuff missing, and it's taking a lot of liberties in terms of stuff that it's just assuming. That might not be true. So, for example, it says, I've already completed but where did I tell it that? I didn't tell it that anyway. It's just kind of assumed it. So if that's not true and you blindly trusted it and you used this email, it could land you into trouble. Well, this stuff about remaining tasks, we haven't told it there are remaining tasks. So it might not be. It's made another assumption. So even though this email is kind of small, it's just full of assumptions. Thing. So contrast it to yourself. Let's just say you are human assistant, and it's your first day on the job and your boss gives you a similar task like this. You wouldn't just go off and do it and make a load of assumptions. You would come up and ask a lot of kind of clarifying questions. Like, if it's your first day on the job and you know nothing, you would instead of going off and writing the email, you would ask your boss, well, you know, what's this project for? What are some of the implications of being late? Does this impact other things, how urgent it is, and a load of other things that allow you to actually complete the task successfully within the context. But whereas if we just followed the raw AI output, it would force us to do a load of extra work. It would kind of mess things up. So the other limitation of these LLMs is called the syncophanty trap. And it basically just means that the AI has this tendency to agree with whatever you say or mirror back what you say to it. You indicate what your preferences are, it's very unlikely to disagree with you. And again, if we go back to how the AI is trained, it becomes a lot more obvious why it does this because it's trained on a huge amount of data. It provides an output, and it's given feedback in terms of whether humans like or dislike its responses. And we're far more likely to like responses that agree with us and confirm versus someone who's disagreeing with us and giving us tough love. So the way these LLMs are trained by nature, they're syncophantic, which for a lot of applications, we really don't want, and it's not desirable behavior. So we have to keep this in mind and use a couple of techniques to overcome this, which we'll go through later in the class. Right, so that's a little bit of background and some of the limitations of AI. Now that we know this, let's move on to the next lesson where we look at the basic structure of what makes a good. 4. Task, Context, Tone, & Organisation: Right, so in this lesson, we're going to build the foundational structure of our prompt, which everything going forward is going to be based off of this. It's the bread and butter of this entire class. And every single prompt going forward should follow this structure of context, tone, and task. So starting with context, like in the last example we saw, we literally just gave it a task to do without giving it any context. And the issue is, as you saw, it will just assume a lot of stuff, and it won't do a very good job at all, and it wasn't useful to us in terms of what the p. And in a way, even though it wasn't really what you meant, it does make sense what it did because it simply just followed your instructions. It just wrote an email with what you gave it with a context. So we must start off every single prompt with a context the situation, any key information it needs, any background. So in the last example, where we give it zero context, so it just filled in the gaps by itself, we can't let it do that. So if we go back to the Eaton where we're asking for a deadline extension, a better context prompt might be something like, you know, I'm working on an engineering project. I've been tasked to do a cost feasibility report. However, it's been delayed by a week because I'm missing some info from key suppliers. They've had supply chain issues and we want to be able to get the quotes back on time. Therefore, I need a one week deadline extension, which shouldn't impact the overall project. And instantly, you get a much better answer without having to tweak the task or other instructions just by giving it more info at the start. So that's context. The next one is tone. Tone is, as it sounds, the way in which you want it to answer to be completely fair, with these LLMs straight out of the box, their default tone is actually very good and has a wide application because by default, they're designed to friendly and yet helpful tone, which applies in a lot of places. But sometimes you might want it to answer things in a more specific way, which is where you would tell it to. So there are a couple of approaches you can take. So the first is telling it the intended targeted audience, so it can tailor in terms of the language it uses for its output. So imagine if you're a schoolteacher, instead of just saying, write me a lesson plan for Henry. Say, can you write me a fun yet educational lesson plan? That's interactive. That's about the life of Henry eighth for a group of year five students. So now it knows who its intended responses for, and it'll tweak it accordingly. So another really powerful tool for the AI is to tell it a situation and also to give it a role that it needs to play. Now, this is actually a very interesting point, and this is a point we'll come back to later in the class because this is a tip that I learned quite a while ago in terms of giving the AI a role. And I've been using this tip for a very, very long time, but the very latest device seems to have tweaked upon this advice a little bit, which we'll delve into in a little bit more because there is some nuance in terms of how you use this. So for now suffice to say, let's say if I'm asking the AI a question, ask it to answer in the way of a certain role. So, for example, I really like to use AI as a reading companion. Let's say if I'm reading a physics book and I come across a really complicated topic, like, let's say, if it's the many words interpretation of quantum physics or something like this, I'll go today and say, you know, answer in the way of a top physics professor and explain to me in layman's terms. And then you immediately notice a difference in how it responds by just giving it that persona instead of just asking it to respond in a default way. And finally, of course, is the task. So the task is something that you can influence the output in a big way by just having a few tweaks and being specific about what you exactly tell it to do. So what it really boils down to is to be specific, evaluate, iterate. So using the example from the official Google documentation, instead of saying something like, write about climate change, you want to say, write a persuasive essay arguing for the implementation of stricter carbon emission regulations. So basically, just the more info you give it and the more specific you are, the better. And now I won't always be super obvious in terms of what you're missing. So this is why you look at the output and then you continually tweak your input to get the best response possible. Now, another little bonus tip when you're giving an AI instructions in terms of doing a task. In sometimes instead of just telling it to do something, it can really help to say to tell it why it's doing it, because a lot of the time, if you tell it why it's doing it and give it more understanding of the background of why it's doing a task, even though you may have missed certain things, it may prompt you. Like, for example, if you're writing some code or a text for something, something like, don't use any ellipses whatsoever in your output response, which technically could work. A better response would be something like your response will be read out loud. By a text to speech. So never use ellipses because the text to speech cannot pronounce them. So once you've given it that context, maybe if you put other characters that aren't related to ellipses, because it knows it's doing a text to speech output, it flag that up. So that's the framework, context, task, tone as a starting point. So before we move on, it's really about understanding what the thinking is to really make a good prompt engineering. So a lot of the time, people hear the term prompt engineering and then they kind of just scoff at it because it doesn't seem that complicated. You're just typing into a textbox, and you just need to write well. Well, it's not really as simple as that. There are a lot of really good writers that aren't good prompt engineering. It's a really specific thought process in terms of how you define a problem and what are the steps to solve it and to iterate. So going back to that really eager assistant example, just imagine if you were talking to a human assistant, you would want to give them as much info as possible to do the best task. The more info you leave out, while still expecting your assistant to do the task, it just forces them to make more assumptions, which exactly is what's happening with AI. So it's about stating things very clearly, having very clear defined communication, and understanding the task. So the rule is to remember never to give the AI a task that a very competent human couldn't do. 5. Honesty Rule vs Persona: Now, this lesson is specifically about the point where we mentioned in the last lesson in terms of giving the AI a fictional role to influence the output responses it gives you. Now, we talked about tone and getting the AI to answer in a way that's more helpful to you. Because traditionally, when I first came across this advice, I think, two or even three years ago, it's always kind of to trick the AI, you know, tell it to play a certain role, give it fictional situation, so it answers in a different way to what it normally does. Now, the main premise behind this tip was almost to trick it to be less lazy with your responses. For example, in the last one, instead of just asking it about to explain this concept, explain the many words interpretation of quantum physics, say, you are a leading physics professor at so and so University, you are my friend. We're having dinner at a fireside chat, blah, blah, blah. You know, you're building up this whole fictional scenario to get it to try and answer the original question. You know, instead of just saying, give me a lesson plan, saying, You are a school teacher. You're in this setting, X Y Z, all of these things. Now, recently, I listened to this podcast between anthropic engineers, saying, with the latest model, actually, this is not great advice because the issue is, when you create these fake scenarios, a lot of the time, it can confuse the AI. Dog it was a screen reader for a microcont We don't know about. I's all right, so I guys like that. That's interesting. Actual into this because one of the most famous things to fight is to tell Billings model that they are some person or some role. Feel like this a little bit better. I see you honest with my mom situation. I may be this experience. Right. Do you think that level of honesty instead, lying to the M or forcing it to I'm going to tip you dollars? Is there one prefer there or what's your intuition? I guess models are more people understand more about the world. I guess I just don't see it as necessary to lie to them. I like lines to models, you know, lying general. But Party, if you're constructing a constructing value set for a machine learning system for a language model. That's very different from constructing a quiz for some children. So people would do things like, I am a teacher trying to figure out questions for quiz. I'm like, M knows what languaals are. Asks can tell you can give you examples they're like, they understand the Internet. So I'm likert tasks I have. So if you're like chest valuation language like, Isn't as I want to do so why would attend you? I want to do some unreleased or titles tasks and I go to someone to work with me and are a teacher, and ois like, Hey, are you? Must were like. And when I heard this, it was, like, a light bulb moment because it was definitely true. What I was seeing was when I was continuing to follow the old advice and to give it all of these fictional scenarios, it was giving, like, these very cheesy responses, and it was wasting a lot of context to play this fictional persona. And it completely makes sense, right? Because in terms of the AI, it's not sure whether you want it to answer the question in the best way possible, or if you're trying to make it to play this persona possible. So then it ends up finding this compromise between, like, these really cheesy answers to play do this role playing game that you're trying to get it to do, as well as answering the question. And basically what the anthropic engineers are saying that the latest models are actually smart enough if you give it the actual true context to understand what's going on and to give you the best answers. So after experimenting a little bit, I do agree with the advice here, because the problem is with the whole persona thing, it was wasting a load of tokens just to play the role and give you a lot of fluff instead of actually answering the question, which contaminates the chat, and it just degrades the overall conversation. And I do think that the best policy is to be honest with the AI. Don't tell it that it's a professor when it's not. Don't tell it it's a school teacher when it's actually not. So having said all of this, does that mean we're completely getting rid of the whole persona thing that we just mentioned in the last so the answer is no. We're not completely getting rid of it because I think there is some use to it, but it just means tweaking the way you give that instruction and having some nuance to it. So instead of saying you are a physics professor, you are this, you know, tell her the exact situation. I'm currently reading this book called Beginning of Infinity by David Deutsch. I want you to act as my AI assistant in terms of understanding the book to bounce ideas off of just to help me gain a better understanding of this I want you to answer in the way of a leading physics professor to explain certain concepts to me in layman's terms. And here is my first question. No, the wording is similar. It's just very slightly different. But it makes a huge difference because now you're not telling it to play this cheesy role. I knows exactly what you want and what you're actually trying to do. Because it knows you're not trying to do this weird role playing thing. Your primary objective here is to understand the concepts in this book. And the AI knows that its primary task is to convey these concepts. In a very clear way. And, the responses are hugely different. You can see, there's a lot less fluff at the beginning at the end. And as always, if you're being accurate about the exact situation you're telling it, if you're being honest with it, it may be able to spot a lot of blind spots that you may not have seen. You know, for example, it may say things like, you know, maybe it's better to learn about this concept as a precursor or maybe this book recommendation that's a different one to the one you're reading. Instead of it being just hyperfixated on playing a cheesy role. So yeah, as a general policy, don't lie to the LLM, I just kind of confuses it and it doesn't help in most cases. Right. So going back to the class project, do we have a foundation or framework in terms of how to structure our prompts? And we've learned about some of the basic limitations of LLMs. Let's see if you can go back and implement what you've learned to our class project. So I want you to use to structure we described, giving it a little bit of context about your background your audience, your niche. Then for the tone, I want to describe how you wanted to answer who your target audience is so it can tailor it to you in detail. And then tell it about the task, giving as many details as you can, instead of just a generic do this answer. And then the most important thing is just to keep iterating. Don't expect to be one and done. Look at the response, iterate it, see if you can tweaks more things, and see what changes that makes. Now, before we move on, another useful tip that you can use control and steer the way that AI responds. I sometimes to tell it how to format things. What you generally want to do with AI is instead of telling it what not to do, tell it what to do. It's slightly more effective. So don't say do not use markdown in your responses. So something like your response should comprise of smoothly flowing pros paragraphs. Or if you find that, as I often do, that the AI is just waffling loads and loads when you just want a succinct answer, don't say, Don't give an unnecessarily wordy response. What you should say is that your answers should be very succinct and gets directly to the point. Answer in the way of an FAQ format, which gives it very clear instructions and less leeway of what to assume again. So that's the foundation. By doing things consistently and in a structured format, you are basically making sure that you're giving the AI all it needs to give you the best response possible. So yeah, give all of those things a try with the class project, and once you're ready, we can move on to some slightly more advanced techniques. 6. Few Shot Prompting: So I want to show you this quick clip from a computer science class that was released by Harvard. But we thought we'd refer to the audience here, and Brian's gonna scribe as we go, and all we want to do this morning is just make a peanut butter and jelly sandwich. One instruction at a time, and each of us will just execute what we hear. How's that sound? Good. Alright, if someone could volunteer with the first instruction, and Brian will type it down. Open bread we heard. Open Bread is the first instruction. So if you'd like to execute, open Bread No don't look at me. Okay. Alright, so we're kind of on our way. Take the knife. But Peel off the cover of the jelly. No covers on ours. Stick knife into the the bottle. From the top. Stick Okay, step. Nine. Rotate hand, so jelly ends up on. Okay. Jelly side down on bread. Poor jelly on bread. All of it. Okay, now you're just messing with us. This rates very well how computers think and the number of built in assumptions we have actually when we give it commands for very seemingly simple tasks that we take for granted actually contain a huge number of assumptions that we don't even think about a lot of the time. You know, it's a funny video, and it shows that even like 20 instructions in, they couldn't really effectively describe how to make a sandwich from scratch. Though the professor and the other students were already overriding some bad instructions. And it's basically showing us that with some tasks, it's so ingrained in us that we don't even think about some of the liberties and assumptions we make. Do get us to break down all of those step explicitly explain it to an entity like a computer that's never made a sandwich before. It's really complicated because we don't have to think about them. Now, of course, it's not exactly the same with LLMs, because LLMs are much smarter than traditional computer programs where you have to give it explicit instructions, and it only does what you tell it to do. LLMs are a bit more context aware and they have some background knowledge with their pre trained data to kind of draw upon. Whereas if you have a traditional programming language, the entire program, could be thousands of lines. It could be correct. But if you put one semicolon out of place, the whole thing might not. And even though the specifics differ, the overall overarching concept is the same in that assumptions are being made. And you could say it's a double edged sword, because with LLMs, they are smarter, whereas the program would still run. It wouldn't stop it doing the task just because you've not specified something properly, like with the traditional programming language, so where they wouldn't run. Of it. But then the other side of it is that might introduce blind spots and assumptions where you didn't expect. So what's a really good way to counteract the LLM just taking too many liberties, assuming things that you might not want? So this leads us to tip number four from Google's five tips for good prompting, and it's called Few-Shot Prompting. Sounds a little bit complicated, but all it means is that a few shop means you're giving the AI a couple of examples of what you want, hence, few shop prompting. So it reduces the level of assumptions that it makes. Samples you give it. The more the AI has to go off of and take fewer liberties. So as an example, let's say if you wanted some recipe ideas, you could give examples of recipes that you already like, can take that into account and output you recommendations based on those things rather than just making guesses at what you may like out of the millions of recipes out there. On AI Studio, you can paste links to recipes that you really like. Just make sure you tick this part where it allows it to go to view outgoing links. Even though you can do this, and it is very convenient, it's not the best way because it'll pick up a lot of irrelevant info that will clog up the context. So the best way is to copy and paste a recipe into text or markdown viles to upload. So going back to the class project, you're a content creator in a specific niche and say you've got five pieces of content that you've already written out, that you really like, you know, with the assistance of AI or you've written it yourself, you really like the content, but you don't know what video title it should have to generate the most click and interest and curiosity. So here is where you want to feed the AI your video script and ask it to generate titles for those videos, that will translate to a high click through rate. But the issue is when you just go in cold like this, as in when you just ask it to generate titles based on the script, but a lot of time it'll just generate very clickbty titles that don't fit in with kind of the whole ethos, of your channel or what you might like to come across. So instead of just saying, generate me video titles for these scripts, what you could do is give you examples from channels that you really like or video or styles or video titles that you think gel really well with you. So, generate me three video titles for this specific video and script. And then here are what I consider good and bad titles. And then you just browse YouTube for good example. So let's say if we're giving good examples from a channel that you might really like, you know, for example, this one, you've likely been playing the game of Life wrong. That's a good example. The world's most important machine, why people are confident when they are wrong? These are all very good titles, and then you give examples of bad titles. So this crazy calendar changed my life. A comprehensive guide to temporal management is another bad example because it's too academic and it just sounds like a boring video. No one wants to click on that. The first three are good because they immediately pique your interest by making you curious. They pose a question. It immediately sows that scene into your brain and you want it to be answered. The bad ones are just overly clickbaity or they're dull and boring. And by giving this to the AI, it has a very clear context of what you want, what you deem good, what you deem bad, and is much better able to align the titles based on your content. And if you combine that with the script, that should massively improve the output of the response. 7. Managing the Context Window: So as we get more advanced into LLMs, our chat is going to start getting bigger and bigger because when you have a really complex task, it's extremely unlikely that you're going to be able to put one prompting in and it be completely done. It's going to be a back and forth ongoing chat, and it's going to increase further and further the size of your chat, the context window. So even though we have our framework of context tone task, it's extremely unlikely that's going to be it. Like, however good your prompt is, you're not going to get everything done by just putting that one prompting in. So which is where we move on to steps four and five, which is to evaluate and iterate. Now, these two steps are very key because LLMs are basically the fastest advancing technology like we've ever seen, even by tech standards. For example, like what used to take a year, it's month to month it's different. Like, every one of the top AI company every month, even on a weekly basis, new features are coming out, and it's extremely difficult to keep up with everything. So which means that if you pick up any sort of tips and tricks and quirks, whatever that works right now may not work in the next month, in the next six months. So you need to have a way to be able to iterate and follow this process and get continual feedback and constantly stay up to date with what's working. So it basically just allows you to continually test and tweak your approach. But like, for example, the fake persona versus honesty isn't a set of tricks that will just stay working indefinitely. It's about constantly keep trying to see what works with the latest models. So it's more of a mindset to have. So always evaluate the outputs and think, like, how can I kind of tweak this to improve a little bit better? Like, what part of it am I not happy about? What parts can I change? And so before we're able to do that, in a very systematic way, we have to be very organized. So this goes back to the whole thing of being a good prompt engineering is not just a good writer, because we have to be very organized, very systematic, so we can iterate feedback, and we can see what works and what doesn't if you're working on a very complex task with, like, a lot of background info or reference info, things like this, you should never type your prompt directly into the chat box. What you should do is have a separate text file to store all of your prompts, all of your input before copying it across. So it doesn't matter if it's a text file or Google Docs, as long as you have somewhere separate to store it, because the first thing it's very hard to keep track of in the textbox. But, secondly, going back to that iterate feedback thing, you need to have a log of things that you tried to as a baseline and then to be able to edit things. Otherwise, you might be just trying the same thing over and over again without knowing. It's just generally good practice, unless you're asking, like, a very casual, quick question. I would always do this. And, of course, I like to date it. So that's with prompts. What if you had more info to give it? So going back to the recipe example, what if you wanted to feed it like 20 examples, either to give you more recommendations on recipes or to take good elements from them? Basically, if you just had loads and notes of info in PDFs Word documents, what's the best way to go about it? This combined with your prompt? Now, technically, there's nothing stopping you from uploading the PDF or the Word document directly into the LLM. I mean, after all, it seems right. It seems correct, right, because after all, in the professional work environment, as an example, PDFs are very widely used. They seem a good file type reliable. But again, it goes back to how LLMs work under the hood. These are pretty much the worst file types in terms of feeding it to the AI, even though they let you do it, because with a PDF, designed to be human readable, but that's not how LLMs work. LLMs process information by taking in raw text. So when you put in a PDF file, it forces the LLM to take all of the text to try and extract and process all of the texts in the PDF. Whenever AI reads a PDF, there's lots of texts split into different columns. It jumbles up all the formatting, and a lot of stuff looks to it like Gibberish. And it can really confuse it. And not only that, it wastes a lot of your context window for it just to figure out what's going on. So as a general rule, with LLMs, when you feed it info, especially with text sort of info, the simpler the file, the better. So something like a raw text file or markdown is like the gold standard what everyone uses. So if it doesn't take you too long, whatever key info you have in your PDF, your Word document, paste it into a text or markdown file, format it correctly, and then you know exactly how it appears in the markdown and text file. That's how the LLM will ingest it. And then that way, you know the AI is just processing the pure information exactly what you're seeing and not just a jumbled mess of text. Now, don't worry if you're not technical and you've never seen a markdown file bef. It's not difficult at all. It's literally just a markdown file with some notations. So, for example, if you put one hash, it'll be a large heading, two hash is a small heading. It's got if you put two stars, it's bold, very simple, formatting things like this, but it's mostly text based. What I'll do to help you out is I'll include a cheat sheet for markdown, and that's all you need. You know, there's not really much learning. You just need to follow the notation. You should be able to be up and running using markdown files. Just make sure you open a text editor and save it as a dot d, and that's literally it. We've talked a little bit about context window. So it's useful to talk about what the maximum length of it is and what the context window actually is. Context window is basically the maximum amount of information that the AI can take into account in one particular chat. The amount of information it takes in is in the form of tokens. Tokens are a little bit like word counts, but it's not exactly word count because not one word doesn't always map onto one token, but just as a very rough guide one average word maps to about 0.75 of a token. So that just gives you a rough idea. So, the bigger and the longer your chat goes on, the more of the context window it takes up. And if you want to find out what the maximum context windows are, you can just look them up. At the time of filming on the free plan for Gemini and ChatGPT, it's around 32,000. So think of it like a working memory. So when you put in your prompt, the longer your prompt, the more of the context window that takes up, but not only your prompt, when sometimes when click on the drop down menu and you see what it's thinking before the response. That takes up context window, and of course, the output takes it up as well, as well as any attachments you put in, which is why I was mentioning about using text files and trying to keep your context window as streamlined as possible without putting any kind of superfluous information in there to clog things up. And then the other really important aspect, we have to take into account don't think of it like fuel for your car, you know, with fuel for your car, it doesn't really matter if it's a full tank or it's a half tank or you're about to run out. The car will perform pretty much exactly the same. It does not work like this for the context window in LLMs. What happens is, the more you fill up the context window, the longer your chat becomes, the more the information degrades over time to the point where if you get a massive context window, it will start to hallucinate. Even though it's within the maximum still technically, it will start to hallucinate. I will struggle to find things. I just won't perform as well. So as you can see on this benchmark, at 128,000 tokens, this model performs at, you know, 84% accuracy. Whereas, if you go to 1 million, it drops down to only 26%, so it's a huge degradation in performance. This is why the context window really needs managing. And another rule is never ever using the same chat, talk about multiple different topics because one, you're using up to context window, and number two, you're really confusing the eye hat. Let's say if I'm talking recipe ideas. And then I ask you about career aspirations and long term goals. That's a really big no no because it confuses the AI and its wasting your context window. So always start a new chat for every specific discrete topic. And even if it's the same topic, if it gets far too long, just start a new chat and summarize what you've talked about. That can really help, especially if you feel that a chats becoming stale and the performance is getting worse. 8. Multi Step Prompting: So we said earlier a couple of times that your AI is basically a little bit like an over eager personal assistant. I that whatever task you give it, it will just run off and do it. But a lot of the time, if it's a more complex task and important task, you don't actually want it to do this. You want to have to slow it down. You know, you don't want your over ego assistant just to run off and try and impress you. You'll be like, Okay, just slow break down the task. You do this thing first. Like, for example, if we take a very extreme example like your house renovation, you don't want to tell your builder or architect or whatever, just took, go and fix up my entire house. Like, that's too generic. That's too wide. There's too much scope. So you want to break it down first. Let's do the floor plan. Let's have a couple of mockups of the interior design. Let's look at the materials. Let's get some quotations. Want to break it down. So you have more overall control of the process and you can guide the AI to do exactly what you want. So this is called multi step prompting. So say we have this task. We're content creator, and we want to email Noon to sponsor our video. So we follow what we learned so far. We give it context, role, and task. So here's the prompt. I'm currently a content creator in finance space. I have 40,000 followers. I use Nian for a long time in my content, and I shared templates with followers. You are my AI assistant in this chat. I want you to help me with securing sponsorship from Noon. Write an email to Nan, asking them to sponsor my channel. To be fair, it's not ad response, if you look at this. It gives you a few options, and as expected. It's a nice email. You see what the issue is? It will just take what you give it, the info about the followers, the template thing, and it's made some assumptions, and it's just sounding very generic and quite obviously written by AI. So we want a few more steps because in this example, we've never written any kind of sponsorship email. So instead of just doing the email, we want to understand a bit more of the strategy behind it. What are the steps leading up to it before we actually send off the email. So we don't want it to do everything in one go, forcing it to slow down by asking it to break down the problem first. So we say, do not write the email first. I want you to follow a few pre steps. Step number one, asked me four clarifying questions which would improve the output of the email so that it is more tailored to give a high chance of success. Step two is to analyze the answers. If there are any follow up questions, then we will brainstorm strategies to best go about it. And in step three, once we have agreed on the strategy, then you can execute the writing of the email. Basically, now you're really slowing it down. You're explicitly asking your over eager assistant to slow down. You need a specific permission that I'm happy with each and every single step before you're able to move onto the next step. Because as we saw in the sandwich example, if we jump too far ahead, it can lead to pretty bad results. Whereas if we can slow down and we can verify every single step this is correct. Now move on to the next step. That can catch a lot of errors. So as you can see, notice actually what we're doing here. We're actually stacking a few of our strategies that we've learned. So, as always, we're starting with our three step foundational structure of the context tone task. We're describing the situation and what sort of output we want. We're asking the AI to do the task over multiple prompts, which naturally it doesn't like to do. It likes to just do everything in one prompt. And the really good thing about this is it really reduces the assumptions because instead of the AI assuming or if there were any brinspots and filling in the gaps, you're asking him to explicitly bring those things up. Multi step prompting is one I use really, really often, actually. I think it's one of the most powerful prompts for the AI. It takes a little bit more time, but because you're splitting your task over multiple prompts, it's almost like a little hack because you're using, like, additional processing power to think about your problem on a much deeper level. I would say, and a tip you can keep in mind is that as I said, like the overeager assistant thing, it might forget and jump ahead, like a few prompts down. So you might ask in the first prompt to say, Okay, we're going to break it down, make sure you get explicit instruction before you move on. And as you start answering some of the questions and discussing things, it might just jump ahead and do the task. So I would say just to remind it every so often, just to say, you know, do not do the task or do not write the email or whatever a task is until I'm happy with things, and then I explicitly give you instructions to make sure it follows. 9. Chain of Thought Prompting: So another variation on technique is called chain of thought. It's another fancy sounding name, but it's actually very simple. Basically, all it is is you're asking the AI to explain its thinking. So when you ask it for a decision on something or its opinion, it will give you the answer and some very high level reasoning as to why it's doing it. But sometimes it can help to ask it to really spell things out. Before giving you the final conclusion. So this would work for things like brainstorming or if you're making a really complex decision. So let's say if you're choosing between two jobs or if you're making big life decisions, moving between places, basically, it's just like a complex decision that doesn't have a clear right or wrong. You want to weigh up all of the pros and cons. This is what you would use that for. Or say if you're a growing YouTuber and you have a set budget you want to spend, and you're not sure if you spend it on, let's say, 10 hours to get an editor to edit your video or a brand new Sony FX three camera. So if you just plug that into the AI, I'm a content creator and then I want to decide what to spend my budget on blah blah blah, it will give you an answer, and it will give you a few bullet points on why each one is good. But then again, it makes a lot of assumptions, and you might not just want surface level reasons that just apply to everyone. And you really want to consider deeply why it's offering any one option and what some of the trade offs so now if we want to use this technique, similar to the f shop prompting, we say, Don't give me an answer straightaway, after you've explained the task and what the dilemma is. I say to the AI, I want you to weigh up the pros and cons of each option and show your thinking for any recommendations. And say it, I want you to explore the implications of either option before coming to the final conclusion. So now it focuses more on the assumptions on the pros and cons, rather than just focus on giving you an answer at the very end. Much clearer to see what goes into making this decision process, and it becomes a much more back and forth process between you and the AI. So then you can say, Ashley, this is quite important to me, but this is not so important. Again, it's about remember to stack techniques. So we have our foundational structure, the context, the tone, the task. Now we have this chain of thought thinking process, and we stack that with a multi step prompt to really get a deeper, more insightful answer. So just a quick clarification because the multi step prompting and the chain of thought prompting seems quite similar, just so that it's really clear. Multi step prompting, it's about breaking down a really complex task into multiple substeps so you keep track of what's going on. Whereas with chain of thought prompting, it's more about having a complex decision where you want to make sure you've weighed up all of the factors that go into it. So that's the difference between these two techniques. 10. Co Pilot Mindset: So in this lesson, we're going to have a little change in pace. So instead of talking about all of the technicals and ways we can prompt better AI, I want to talk about the mindset behind using these models. So by this point, we've got a pretty robust system in terms of squeezing performance out of the LLM. By now, we can see that we can't treat AI as this all knowing entity that knows better than you in everything. And you have to prompt it in a very specific way to get the most out of them because they do have their own individual quirks. So, for example, it's a really interesting conversation from DemisHsabs, the Deep Mind Google founder. And at least at the time of filming, anyway, he says that these AI LLM models have these really jagged areas of intelligence. And what he means by that is, like, in some areas like mathematics, as an example, it has, like, basically PhD level knowledge. But also, at the same time, when you ask you to do, like, very extremely simple things like count apples or count fingers and stuff like this, it gets it horrendously wrong. And not only that, it gets it wrong, like, confidently, you know, any elementary school child can do. It's less like having this all knowing entity in your pocket or on your laptop, but more like an extremely intelligent and overeager human assistant, where it can help you so much, but you can't just switch off and you have to maintain oversight in terms of everything it does, and it takes a little bit of management. In a way, I think it's very cool because it takes the cognitive load off of you in terms of doing all of the boring tasks. But it doesn't mean you can just switch off and not think. It's just now you direct your thinking in a different way. So as an example for a content creator, you can use it to generate ideas and titles and things like this, but it still requires you to look over it and to use your human judgment in terms of what resonates with you, what goes with your channel what really you want to talk about. So it has to be in an area where you know something about it to double check it. You would never use AI for something you know absolutely nothing about and to do a task that you can't check because it doesn't work like that. It could be absolutely whatever it is. If you're writing an email, you know, you need to be able to proofread it to make sure you know, it lines up. If you're using it for legal work need to manually check all of the case law. You need to check its arguments are actually robust. It can do a lot of it for you, but it does require you to verify things, and you would never use it just to write code, and never look at it and just hope it runs and everything works exactly as you expect it to. So yes, remembering that rule of not giving a task that a human couldn't do, but also once it does that task for you to verify everything that's done, and you're happy with it. So this is why I call this lesson the copilot mindset. You're still the captain and it's a super useful entity to have. But it's about really understanding how to best work with it going forward. 11. Red Teaming: Now, good. Now, so going back to our foundation where we talked about iteration, this red teaming is a really important step, particularly if you're doing a really complex task. So, you know, we've done our usual prompt, the tone, the context, the task, all of that really good stuff. And we're going back and forth. We're doing multi step prompting, and it's all going very well. Like, the AI is agreeing with what we say. Remember, early in the class when we talked about the whole SincoFancy trap, this is the issue here because we can follow our prompts exactly as we outlined. You know, we've got a multi step prompt. It's agreeing and we think we're converging to a solution. But the problem is, as we said, AI, by its very nature is designed to agree with you and to mirror what you say. So as a way to verify and overcome that is to red team it. So to put it simply, we want to prompt the AI in such a way that it thinks we're acting on the other side, and our preferences are the opposite to what we actually want. So it gives us other opinion. So it's not biased just to agree with us. So even though we did say, in general, it's a good practice to always be honest with the AI, I feel like this is one of the few exceptions where you would try to trick it because it's such an ingrained behavior of AI and how it's trained. I mean, to be completely fair, with the latest models, because this issue has been going on for so long and there's literally memes and stuff about they have improved it. Like, the latest AI models are designed to push back against you very slightly. But in my opinion, I just don't think it's enough. It will just default back to its old behavior. So let's just say as an example, you're going through a court case. You know, you think someone owes you money. You give it all of the evidence, you have a big chat, you go through all of the techniques. And the AI saying, you know, yeah, you have a really good case. It's really strong. The evidence is in your favor X Y Z, and, you know, you're feeling really good about yourself. What you really should do to get an objective view is start a new chat specifically, so it's not contaminated. Start a new chat, but pretend you're on the other side on the defense with the exact same evidence and ask it, how strong is my defense? And then that would give you a much clearer indication, because if it's agreeing with you both sides, that's not very good. But even though it's not very good, re pick the evidence, it will cherry pick the evidence that supports what it thinks you want it to say. So it will cherry pick like, lots of good evidence for the defense. And the prosecution. And then you can independently weigh up those things and have a much more well rounded view rather if it was just one side and it almost gets to the point where it's kind of just gaslighting you based on what you want to hear. So this is actually really important, particularly if it's a big decision if you're asking for kind of verification, validation on, you know, like, a big thing, like, for example, you know, a legal case or even a career change or life decision or or whether or just anything where it's like, really nuanced and you're looking for some validation because I'm not even joking. I've literally had it where the AI just completely flipped its conclusion, like four or five times, literally in one conversation, and it's like, really infuriating. You know, you say, Okay, so this is what I think, based on all of the evidence and based on all of the context, everything I give you, like, do you agree? And it's like, Oh, that's a great idea. It's really insightful. And then I'll say, actually, no, I thought about it again. I don't think it's the best. And then it will say, that's the most insightful thing you've said in this whole conversation. This option is better. And it'll keep seesawing, and it can get to the point where it's actually ridiculous. It is getting better, but it does still do this sometimes. So what I would do is read me and start a completely new chat with each chat saying, you know, with each chat, indicate a AI, you have a preference in the opposite direction in the separate chats, and then see what it says, see see how much it validates each of those opinions and see what points it gives. And then you would have to kind of independently weigh those rather in addition to what it's telling you in the chat. And then that gives you a much more well rounded picture. So yeah, this one is definitely what I would do to overcome the whole sycaFancy thing. 12. Automating Context (Custom Instruction): Right. So I've thrown a lot of techniques at you in this class. And as you experiment with these techniques, and as you start using them in everyday life in different applications, you know, you do your whole foundational structure, you do, I don't know, your multi step prompting or whatever it is, you notice that you're writing the same commands again and again, like, in terms of, like, certain instructional cues or like, to or you might notice it's doing certain things a lot and you're telling it not to do a certain thing, again and again. So it's here where we really make use of using the automated Custom Instructions that come with LLMs. So it remembers what your preferences are. And so I've saved this lesson until quite late in this class, because by now you know how to iterate your prompts, you know how to do the foundational thing. And you've generally seen how AI works and the reasoning behind them. So you can set your custom instructions to your preference even if you take custom instructions from either myself or somewhere you've seen online, at least you know the reasoning behind them. And the reason I've left it so late is because even though technically, you know, there's nothing binding about them, you can always change them. What I found is certainly for me, anyway, once I put in my instructions at the very beginning, it's something I almost, like, forget about and don't really come back and actively change. So I think it's actually quite important to have an understanding and set, quite good Custom Instructions from the beginning, and then that will serve you really well going forward. So to access these, they're quite simple. On Gemini, it's like settings, instruction it very similar on ChatGPT and Claude. It's just somewhere in the settings menu and personalization. So when you set these, think of these as like an invisible layer that sits in the background of your account. So it's a permanent set of, like, overarching guidelines and principles you want it to follow for every single chat before it even looks at your prompt. Straight out of the box, LLMs are designed to, of course, have the widest possible applications. So they're kind of designed to be chatty, helpful polite assistance. And we'll just make really big assumptions on what it doesn't know, just so it maintains that maximum level of helpfulness and just fill in the gaps by itself. The big, big issue with this whole persona of being a really helpful assistant is the whole sincaFancy thing. It just basically agrees and mirrors what you say. So LLMs out of the box are like a very good t shirt or dress. You know, it probably fits most people. But however good it is, it won't be as good as something that's tailored, which is what we're trying to do here. So a lot of what you put here will be personal. I'd recommend you starting to use the techniques in this class, experiment a lot and settle on the method that you like. And then naturally, you'll find that there are certain things you keep telling it to do over and over again. And you can put Custom Instructions here. But anyway, what I've done here is I've given you some examples of ones that have really worked for me. So number one is the sycophancy r so the anti syncancy rule is the one that has always been an issue. So I have a rule here to force it to give other points of view. So I say prioritize objective truth over agreeing with me. If I present a strategy, idea or argument that's flawed, do not validate it. Actively play devil's advocate, point out the weakest links in my logic, and highlight disconfirming evidence, even if I might not want to hear it. So as we know, by nature, LLMs are they act syncophantic or agreeable, like, no matter what you do. But this forces it to give some points to the contrary, and hopefully it'll bring stuff to your attention, and you can ask it to explore more if it's an issue. Which could help with reducing your blind spots. Number two is checking assumptions. One of the biggest limitations of LLMs is that if you have a very vague point or even a very well fleshed out prompt on a really complex topic that might be missing some context, it will just guess and hallucinate an answer in the very worst cases. Basically, it's designed to say, what if answer that probabilistically seems most plausible to be the answer with what it's got. But if you don't give enough info, it's going to be a very bad answer. So we need to force it to slow down and ask clarifying questions. Never guess my intent or make assumptions if a prompt is vague. Lacks constraints, or is missing a key context. Instead, stop and ask me a bulleted list of clarifying questions before you generate any response. So in my experience, this one in particular is hit or miss because the AI does not naturally like to defer doing the task over multiple messages. It just wants to do everything in one message. So going back to what we were saying about the overly eager assistant, it just wants to do it. So it's helpful to have in here, but you have to realize that you may have to just manually tell it to do it sometimes. Number three, the fluff and flattery filter. So I don't know why, but this one bugs me more than it should. Sometimes it's actually really infuriating, especially if I do my book reviews or whatever it is, when LLMs, flatter you with just insincere bull crap, for example, I'll be doing a book review with AI asking it certain questions that come up. And it will say, like, really over the top things. Like, that's a really extremely insightful question that gets to the core of the issue. Or they'll say, like, stuff like this shows you really thinking three steps ahead. Things like a human would never say to, like, just a basic question. Plus, it adds a lot of fluff. That's just a waste of time. And it infuriates me. So I say, skip all conversational feeler Itancy of flattery, robotic intros or outtros. Never start a response with certainly I can help. Focus purely on substance and get straight to the direct answer. So, I like to use this quote. This one works quite well as it's a direct instruction to the AI, as an added bonus to it, not being annoying. It saves you extra tokens for more useful stuff or clogging up the chat. Number four, intent focus. Sometimes we write prompts quickly and we don't use the exact words you want the AI, and you want the AI to read between the lines. You know the thing that happens when particularly you're in a casual conversation with your friends and your words don't make sense. But they just go, Yeah, we know what you mean. Kind of like that for the AI, if that makes sense. So focus on my implied underlying goal rather than a strict literal reading of my prompt. Adapt your response to solve my actual problem. If my or intent is ambiguous, explicitly ask me to clarify. So these are some of the simple instructions that I use to help AI slightly overcome some of its limitations and to force it to do things. That's more tailored to me. So it's basically tuning the AI to be, a little bit sharper and take less liberties. Because they just work in a background, you can literally just set it and forget it. 13. Final lesson: You've made it to the final lesson of the class. So statistically, 87% of students who start a class and do the first lesson don't make it to the final lesson. You're in a very select group, and as a reward, I've got some resources for you. Because we've gone through a lot of techniques, and I don't want you to feel overwhelmed. So this video is just to round things all up. And going forward, it will be a continual process of experimentation. Just from the nature of LLMs and how fast they're advancing, it's literally impossible for you to have a set of framework or tricks or tips that will continue to work indefinitely. It's having that method, that mindset of thinking constantly iterating, constantly improving your responses. So I've got some cheat sheets just to summarize everything we've talked about what different techniques there are, what are the best way to improve them, and everything just to kind of jog your memory and to give you a little bit of inspiration when you're working on a really tough problem. So if you want to stay in touch with future classes, future developments of AI, and things I'm reading, just general interests that I have, I actually publish a weekly newsletter, so make sure you sign up to that, and you have all the latest updates on things that I'm working on, as well as quotes and other resources based on investing. So also I'll be very curious to see in terms of the class project. That has evolved over the multiple prompts. So if you're happy to, you know, feel free to share your prompts. That's worked really well for you and how that's changed the output of your responses, so other students can learn, can chime in. That'll be really useful. It's a really, really exciting time because this technology is the fastest growing, most exciting technology, in my opinion, anyway, that I've seen in my lifetime, and getting really good and mastering it will really give you an advantage going forward in the next, you know, 25 years. And it's a process of experimenting and just having fun with it and trying out new things. There are so many applications, I feel like people are just starting to scratch the surface now. It's still really, really early. So have fun with your class project, post the output of your content and stay in touch and enjoy. Thank you very much.