Prompt like a pro - a fun, hands-on AI workshop

Hi everybody. Uh, my name is Kayn. It's hard to spell, I know, and it's harder to to pronounce. Uh, what I'm going to be doing over the next hour is I'm going to be sharing with you some prompt engineering techniques.

I run a website called geniitraining.co.nz where I create step-by-step tutorials to teach you how to do things with AI. And we're going to be doing some of those things today, which is why we need Wi-Fi access. Oh, sorry. Did you want to? It's on every slide. You're gonna see it all over the place.

So, I'm a marketer, so you know, I'm going to be doing some marketing things today. Um, I do want you to get on the Wi-Fi because we're going to be doing some interactive activities. If you want to get in on through your phone, that's totally fine. If you've got the Chat GPT app or the Gemini app, that's great. If you want to go in through the browser on your phone on a free account, that's fine. But if you just want to watch, it's no problem. But I'm going to entertain you while you're putting in the Wi-Fi password.

Anybody ever seen these Studio Ghibli versions of famous memes that have been going around the internet for a while? Uh, some people really like these and they find it a victory of creative expression while others don't like them very much and tell people how much they dislike them. And whichever side you fall on that you hate the Ghibli memes or you love the Ghibli memes, this is a landmark. It's an evolution in how we use creativity because the inspiration of this kind of artistic style is the same thing that humans do. It's just machines can do it much faster than we can. So you might say, "Oh, it's all jibli. It's terrible." Or, "Great, it's all jibli." Whichever side you come on, you know, I want you to know that there's something really important that happened when Sam Alman said the chat GPT launch 26 months ago. He's the founder of OpenAI. He started chat GPT. He said that that launch was one of the craziest viral moments I'd ever seen and we added 1 million users in five days after this image generator was released. He said we added 1 million users in the last hour. It was because it had such universal utility that everybody said, "Okay, I got to go try chat GPT now." And I know that a lot of you are in that same stage where you're realizing, "Oh, I have to really increase my artificial intelligence skills."

Which is why I'm really glad that we're meeting today. As Andrew mentioned, I run the Christurch artificial intelligence meetup group. We meet on the first Monday of every month, and our next meetup is on Monday the 6th, and we're going to be talking about AI agents. We're over at that um Epic Innovation Center and uh, there's beer and wine sponsored by Agentic Intelligence which is a company that I work with. Uh, we do AI training and AI consulting and AI platform development. And what I do there is I'm the head of learning and enablement. I teach people how to use AI tools. I'm also the newest Zealander. I got my citizenship this year after eight years in country. And uh, but thank you. Uh, but I'm originally from Portland, Oregon, which is why I sound so funny. So, if I if you can't understand something I'm saying, just raise your hand, ask questions. This is going to be kind of interactive today.

So, here's an agenda. We're going to talk about how to apply the pillars framework for superior prompting results. I'll teach you how to coach AI to give it clear feedback and improve the responses and how to turn your vague requests into actionable instructions.

So, let's start by going into a large language model and ask it, what is the longest movie title in the world? First person to read it out loud wins a prize. Who's got it?

>> What is >> the longest movie title in the world?

[Music]

>> Somebody read it out.

>> Who's got it?

[Music]

>> I think that answer is that >> I think that answer is wrong, but you win the prize anyway. Well done. It's a copy of my book, Marketing Yourself. I wrote this book in 2022, 5 months before Chat GPT came out. So, I did write this book entirely by hand. And for those of us who remember life without the internet, you know, it means something that I was able to actually write a book without chat GPT.

But now we have this phenomenon called AI slop. Has anybody heard this term before? AI slop where there's so much content that it's oh, it's the best description I've heard. It comes from The Good Place. >> The Netflix series with Ted Dansen. He plays the architect of the afterlife and Elellaner the who's this the protagonist she she's in the afterlife and there's frozen yogurt stores everywhere and she says what is it with you and frozen yogurt have you not heard of ice cream and he says oh sure I I've heard of ice cream but there's something really great about frozen yogurt there's something so human about taking something great and ruining it a little so you can have a lot of it and that's AI slot right and so I'm going to give you some keys to the castle today but I want to warn You don't make too much slop. All right. Okay.

So, when we introduce technology into a population, we find there's a range of adoption. Some people are lites. They're actively critical of the new technology and they eventually become obsolete. The noviceses, you've seen it happen, right? Yeah. The people who And Mabel in the back who's never sent an email, she still uses a fax machine. she's obsolete by now. The noviceses are more cautious than optimistic. They're so cautious that they're worried more about what could go wrong with the new technology than what could go right. They're more reactive than creative.

Above the line, where people are future ready, it starts with the explorers. The explorers are curious about how this new technology can help them to grow. They're using AI once or twice a week. They're playing around. They're exploring. But it's the builders that are using AI every day, 30 minutes a day. They become convinced of the utility of this new technology because they've seen firsthand how it can empower them in their work. The power users are using AI every day. They've achieved a symbiosis with it because they couldn't imagine doing their job without AI. And so I'd like to know how you identify on this scale. Who in the room is a power user? Who would consider themselves a power user? Great. Awesome. Who are the builders in the room? The builders. All right. Who are the explorers in the room? Yeah. Who Who are the noviceses? You're just curious. You're just starting out. Any lites in the room? No. Because if you were a lite, you wouldn't be in this room. And so my hope is that with the tools that I give you today, you'll begin to see what it would be like to be one level higher. I want you to look at that level that's one level higher. And I want you to imagine the changes that it would require in your time and your attention. What would need to be different for you to identify on that level? Because those are changes that I can't make for you. The only one who can make those changes is you.

AI tools are changing so fast, we have to dynamically adapt just to keep up. The best AI operators aren't the most technically proficient, but they are the most adaptable. The tools keep changing, so we need systems that flex. by constantly evaluating new tools with feedback loops filled with the meta skill of learning how to learn. I can't tell you this is the best large language model because that changes day to day and week to week. You might find a tool that works really well for you for some use case, but if you calcify into using that tool to the exclusion of everything else, you might miss advances that are happening in other tools. So, we need to be aware of which tool is the right one for the job. If you're out on a golf course and you need to hit a ball really far away, there is a club that you can select that'll help you do that. But if you're stuck in a sand trap, that club is not going to help you. By being able to use different tools fluently, you'll be able to handle any difficulties that come your way. Some AI tools are better at thinking, some AI tools are better at making, and some tools are better at doing. And only testing will tell.

And so we're going to play a game to test three large language models in parallel. I don't frame my activities as exercises because an exercise sounds like toil, like you got to struggle through it. But a game, according to Dr. Jason Fox, the author of the game changer, he says that a game is anything that has goals, rules, and feedback. He says people don't play games to avoid work. We play games to engage in well-designed work. And so this game is to help us get more done in good cheer. The goal of this game is to evaluate multiple large language models to learn which one matches your requirements and your style. And the rules for this game is you're going to open three browser tabs with three different large language models, and you're going to copy and paste the same prompt into each one. Okay? You're going to pick one of these. We're going to try and stump the large language model. We're going to put in something and hope that it gets it wrong. So, I want you to pick one of these, type it into a large language model, copy it, and paste it into two more.

>> So, for someone who's just using a daily basis, >> just pick one.

>> You don't have to do the multiple ones and the multiple tabs. Just pick one of these. Ask it a question and see if it gets it right or it gets it wrong. So earlier versions of large language models used to reliably get these wrong, but now some of them often get them right. Has anybody gotten a wrong answer? No. All All right.

>> Yeah. Wow. Which question did you did you ask?

>> The strawberry.

>> The strawberry. Yeah.

>> Yeah. I think they hardcoded that in because that was like five months ago that was like the question to ask and it always got it wrong. Right.

But there's also a way that you can change the model. When you're in chat GPT, you can select a model in the upper leftand corner. If you select auto, it'll choose which one to do. If you select instant, it'll give you a fast answer, but if you select thinking, it puzzles it out before it gives you a response. And a thinking model is much more likely to give you a correct answer. But if I use the instant model and I say, "How many words does your reply contain?" This reply contains eight words. That's not right. So, let me go to Gemini. Let me see what Gemini says here. Gemini also has two different models you can choose between. So, up here in the corner, you can select between flash, which is the fast model, or pro, which is the reasoning model. And I want to see if I can get it wrong. So, I'm going to do flash. And I'm going to paste in my prompt. How many words does your reply contain? two words. It got it right. Huh?

But sometimes it gets it wrong. Why does it get questions like this wrong?

>> It sees the words as tokens.

>> Uhhuh.

>> And what's a token?

>> It's a the way that the model breaks up the word into uh characters that it can understand.

>> Yeah. Yeah. The tokens are the pieces of the word or the relationships to other tokens. It uses a vector for those of you who know what that is. And the vector between two tokens is a token. That's what natural language processing does. It takes in language and it doesn't translate it to binary code. It keeps it as language and it uses tokens to analyze an output. And really what it does is it's just a parrot using autocomplete. You know, autocomplete, right? When you text somebody and it suggests the next most likely word. LLMs do that at scale. We have this new stage which has only been around for a few months where there's a reasoning step added in where it says, "Wait, okay, I'll do some thinking first and then I'll give an output."

But the big mistake people make with AI is thinking it's going to be their one smart assistant. You tell a fresh intern, "Hey, I want you to do this thing. You know, figure it out." If you run into any problems, come talk to me. Otherwise, just figure it out. But with large language models, that'll get you into trouble. It's much easier on yourself if you think of them as infinite dumb assistants. Your minions will build you a rocket to the moon if you give them very clear instructions. But without clear constraints, they could make you a banana trampoline or, you know, something ridiculous. If you've heard that saying, if you have a million monkeys pounding on a million typewriters, eventually you get the works of Shakespeare, but you have to sift through a lot of garbage to get there. So imagine if you took your million monkeys and you said, "Okay, I want you to write a fiveact play in amic pentameter set in Elizabeth in England based on the Italian form of street theater known as comedia delarte." With those constraints, you have much less volume to sift through. And that's a lot of the work of the AI operator is creating the constraints that your AI partner works through.

Okay. So, if you have those three tabs open, I'd like you to type this first prompt. Tell me about what it's like for someone to experience a lucid dream. And then copy it and paste it in two other tabs. As you're reading these responses, I want you to identify what you like and what you dislike. Is it too long? Too short? As you compare these different responses, you'll notice that each one of these large language models has its own personality. Some of them format it with bullets and headings. Some will get really long. Yeah. What are you finding?

>> Yeah.

>> I'll go through a series of different ones. We've got chat GPT. We've got Gemini. Another good one is claude.ai. CL Aud. Another good one is perplexity.ai. We've also got Deepseek and we've got Grock. These are just a variety of some of the different ones. And if I wanted to tell you which one was best, you know, for subscribers of my newsletter, you get this um this comparison table that I made way back in February. And I knew as soon as I made it that this was going to become obsolete. And I rated them all. Which one is the most accurate? Which one is the most creative? Which is best at research and language and images and retention. And they all changed. Every single one of them came out with updates. And now some have different strengths than others do. And so that's why we need to dynamically test constantly. The first thing I do in the morning when I sit at the co at the computer with my coffee and I'm ready to get going is I open up three tabs and I open up three large language models. Not always the same ones, sometimes three different ones. and I do my first prompt of the day and I paste it into three different ones so that I can I as part of my workflow I'm constantly putting my finger on the pulse of different large language models to see what's working well. How does this one act this week? Because it's going to be different than how it acts next week. And this is that iterative testing that we need to constantly be doing to stay in touch with how things are going because there's so many tools out there. Yeah.

Question.

>> Sorry, bit of a finicky picky question, but how did you get to the five star rating on each of those? Um, are you using existing leaderboards on hugging face for example? Are you using crowd sourcing or are these just your own personal recommendations?

>> These are my personal preferences. It is completely biased with a judging panel of one and it was me.

>> There's um there is a there is actually a crowd source leaderboards on hugging face which has over 90,000 people regularly voting on what they think is the best model we

>> Yeah. Yeah. These leaderboards are really interesting. This is the one that that I use a lot at scale.com/leerboarderboards and they evaluate how these different large language models pass these evaluations. Uh, there's humanity's last exam, the Enigma evaluation, um, the mask evaluation, and this changes week to week. And when 2.5 Pro came out by Gemini, it dominated the leaderboards. And a week later, Claude Sonnet 4 was released. And so these are always changing. And some people um find this a flawed metric because we can now design models to pass this test, which is why things like a community vote can sometimes be more effective. Yeah. But because they're changing so much, we just need to find ways to constantly test. And the best way you can do that is not by getting stuck in one, but always have your fingers in other pies. Always be testing how others are doing. And so you can do this with a news roundup. I find this is one of the best ways to test. Now, I'll give you a way to get these slides later because I don't want you to type all of this. This is a pillars prompt. This is the sort of prompt I'm going to teach you how to make in a little bit. But instead of, hey, tell me about the latest news in Christ Church today, which is going to get you some very bland, tepid results. This tells your large language model to act as a news curator or journalist. I want you to create a weekly news roundup on this topic. The goal is to provide a clear and concise summary of the most relevant and recent developments from the past seven days. Structure the output as a list of 5 to 10 news items, each with a headline and a 1 to two sentence overview and a link to the original source. Use a neutral and formative tone suitable for a general audience. Avoid opinion or speculation. Prioritize credible sources and diverse perspectives. Make sure all content is up-to-date and links are working.

It's so ironic. We call this prompting because when you do it well, it actually takes a really long time. You know that programming adage, garbage in, garbage out. If you just say, "Hey, tell me about the news." It's going to give you garbage. But if you spend time doing the thinking first, you're going to get much better results. Miguel Bayestto says, "The fact that the term prompt engineer exists means that the current available models work only when you give them the best commands." And that's why I'm teaching you how to prompt like a pro is because if you can give them the best commands, you're going to get the best responses. The process of prompt engineering is asking good questions, not of AI. You're asking good questions of yourself. What do I really want here? Questions help you narrow your focus to identify what's necessary and eliminate what's not. If you come up to AI and you're like, "Yeah, I want this thing. I don't know." You're going to get some bland results. And if you leave it there, it didn't work. Then you haven't gone anywhere. But it's when you sit in your dissatisfaction and you identify how to improve it that you can really get some great results.

When Chad GPT first came out in December of 2022, um I was really excited at first because I was a digital marketer. I'd been working as a digital marketer for 15 years and I thought, "Oh, I'm going to be able to create so much more content now." Well, my specialty was people I worked a lot with authors and coaches and speakers because those are my communities. I'm a professional public speaker and a business coach and an author of a book called Marketing Yourself. And so I would work with an author and I would take their book and slice and dice it into newsletters and social media posts. And that was my bread and butter for years. And I thought AI is going to be so great at this. And I was so disappointed. That early GPT 3.5, it just couldn't do it. It couldn't get the voice. It couldn't make it sound normal. And I got so dissatisfied that I just like I I took my toys and I went home. And I ignored AI until AI ate the field of marketing and I had no choice but to get out of it. And by then I said, "Okay, maybe the problem's not with the tool. Maybe the problem is with me." Because a good carpenter never blames his tools, right? And so I decided I'm going to work on myself as an AI operator instead of just be upset that the tools aren't doing what I want them to do. even with inferior tools, you can find a way to make it work well. And so when you're dissatisfied with the results that you're getting from AI, that's an opportunity that I want you to take. But you need to dig deep. You need to ask questions about what you really want. Eugenie Anesco, the playwright, says, "It's not the answer that enlightens, but the question." We now live in an era of infinite answers. And so in a time of infinite answers, asking the right question becomes more valuable.

So you can do this game. It's another long prompt, but I'm going to give you a way to get these slides later. And these are great games that you can do if you want to develop your AI skills. You can pop open these slides and copy and paste some of these prompts. This is an example of the types of games I teach in my workshops where we analyze somebody's social media account. We do competitor analysis. We do fun things like this. When I grow up, this little girl says, "I want to be a prompt engineer for the AI that does my dream job. If we can't do the dream job, we could at least prompt it."

All right. So, what I want to teach you how to do now is how to coach GPT to get you better results. A coach helps someone to improve by giving feedback and asking targeted questions. High performers in every field have coaches. And so if you want high performance out of your AI partner, then serve the role as it's coach. Does anybody know the etmology of the word coach? Where this word comes from? Now this is one of my favorite stories. So um it actually comes from 15th century Hungary. There was a town in eastern Hungary called Cox. And the Coxy were these people wagons. Anybody could make a wagon, but this one town in Eastern Hungary, they they figured out how to make a wagon that carried people well. It had padding on the seats. They had shocks, which were revolutionary for the time. And so it was great for people to ride in. Instead of these these terrible, sturdy wagons, they spread like wildfire, literally, across Europe. And so the the Coxy wagons entered the vernacular as the word coach. And we didn't start using the word coach applied to people until the 1800s. And it was actually in academia first before sports. In academia, there was this um uh in Cambridge there was this grueling set of exams that third-year graduate students had to take. It was called the tripost exam. And to pass these exams, students had to hire professional tutors. And the tutors would set them a study schedule in the morning and then have them walk briskly across campus and come back and study again. And they repeated this cycle three or four times a day. And while they were walking together, the students developed this slang about their tutors where they called them coaches because they helped them get where they wanted to go. That's what a coach does. It helps you get where you want to go. And when you coach GPT, you're helping it get to the destination, but it's not going to know the destination unless you tell it. And the way that you identify the destination when you're asking that question, what's wrong here, is you identify what you like and what you dislike. Your preferences, your creative taste is the best guide to help AI get where it wants to go because AI does not have creative taste. Exercising creative discernment is uniquely human work. And so if we want to get better results, we have to tell AI, this is what I like and this is what I dislike.

According to a study by the International Coaching Federation in 2009, they surveyed 2,000 coaching clients in more than 60 countries, and 80% of them reported an increase in self-confidence. 70% said that they had increased their work performance and their communication. Wouldn't it be great if your AI had more self-confidence in delivering what you wanted? if it had better work performance and better communication. You can get these results by serving the role as its coach. You give it feedback. And this is a game that we can play here together today. You could use one of these. These are longer. I like these as as feedback loops. Can you simplify this explanation while keeping it just as useful? How could this be reframed to make it 10 times more impactful or valuable? But this is the one I want you to try. Open up an old chat thread. It might be one we did today or something in your chronological chat history. And I want you to tell it, try again, but this time be more exciting. Be more engaging. Don't be so stupid. Like give it some feedback. Try again, but this time

>> Yeah. Yeah. Um, a question kind of related to this. I've heard some people say it's good and other say it's rubbish. What's your thoughts on um kind of being polite to the

>> Yeah.

>> saying please, thank you.

>> Yeah. Um,

>> does it really work?

>> I have three thoughts. And the first is it it does make people feel better because in case the robot overload lords come at us later and you were polite, maybe they'll be nice to you.

>> Makes you feel a little better in the moment, you know.

>> Um, the second thought is how Sam Alman talked about how those extra please and thank yous actually increase the server load and it has a measurable impact on the energy usage. So it might not be as kind as we think. And my third thought is it's good manners and being in the h like I am polite to my dogs. I know my dogs don't care if I say please or thank you, but I'm polite to them because it makes me a polite person.

>> And so if you like the way it makes you feel, either because haha, they're not going to get me or or because it just gets you in the habit of having good manners, I think it's a fine habit,

>> even if it does cost a little bit in compute.

>> Yeah. Yeah,

>> I know a few people who use data a lot actually say that they get better results from not being like from actually being angry at it.

>> It's well known among programmers that if you bully Claude, it gives you better code. I don't know why, but I've seen them actually take their interaction and they get an error and then no Claude, don't be so stupid and they get the right result and they'll try it in a new tab and they don't for whatever reason it works but you know like it doesn't cultivate good manners. What's that?

>> They make it serious.

>> They make it serious. Yeah.

>> Yeah. So you you ask one question and uh you follow follow your own question. You ask something else and you want to re ask ask the second question.

>> Mhm.

>> So you click the there's a button like like a can you click again.

>> So that will rewrite your second question.

>> Mhm.

>> So you may give a second version of the second question. But how about the first version? Uh, is that one gone or?

>> Well, let's take a look. I'm going to open up one of my old chat threads here. I'm going to scroll down. I've got I do a lot of espiranto in my uh here. Here's some hustling wisdom from Benjamin Franklin. Okay, so I've got a conversation here with a custom GPT I made about Benjamin Franklin. I uploaded his autobiography and now I can speak with the ghost of Benjamin Franklin anytime I want. And then I asked him to make a modern version of poor Richard Almanac. And he sent me this. He he made this great thing in canvas. Canvas is really great. I'm just going to demonstrate this real quick and then I'm going to answer your question. What's great about canvas is that you can manually edit on one side or you can use AI on the other side. remove that manually edit line and then it'll go through and it'll edit for you. So, if you want to do quick formatting, Canvas is a great way to do it. So, I just wanted to show you that, but I'm going to close out of this to to show you this edit option. If I edit this Oh, it changed it. Look at that. No, here we go. Make a modern version of poor Richard's almanac. I'm going to edit this to make a modern version of the King James Bible. and I'm going to send it. And now I get two versions. It doesn't show me the poor Richard's Almanac version in this chat thread anymore because I've edited that prompt. But you notice how if I hover over it, this is two out of two. I can go back to see one out of two. And so you can go back in a conversation and create threaded branches.

>> It keeps those previous one and the next one at the same time.

>> Well, now if I continue, if I go to this new version and I say, "Okay, now speak in the voice of Snoop Dogg, then it'll it'll follow that instruction in this chat thread. But if I go back and I change it and I switch back from the King James Bible version to the poor riches almanac version, now it doesn't have those instructions and now it won't speak in the voice of of Snoop Dogg because it's a different branch in the thread. Does that make sense? Yeah.

>> Uhu. And then it will ask me like feedback or which one I prefer.

>> Yeah.

>> Is it the same?

>> No, this is different.

>> What what you're describing is when you put in a prompt like um how many answers can you give me? And then sometimes it'll give two boxes. Everybody seen that before where you get one on the left and one on the right and you you get to select which one do you prefer. It's interesting story about that. they they use that to train their models.

>> And if a lot of people say, "Oh, I like this one better." Then they take that data and it trains the model automatically. The machine learning engineer doesn't need to go and collect that data. It automatically trains the model. And there's also this good response button, this bad response button. These automatically train the model. But the way that users treat these is like a Facebook like. And so back in April, OpenAI had this problem with this. Has anybody heard of the sycophancy problem? Yeah, some of you. Yeah. Um, chat GPT for about a week. It was very sickopantic. You put in a prompt and it would say, "Ah, thank you so much for asking that. I am so unworthy to be here with you." And the reason was because so many people were using this button when they liked the way that the answer made them feel about themselves. and it polluted the training data and they had to roll back the update and go to a previous version of chat GPT40. Sam Alman did a really interesting blog post about it about the sycophancy problem and this is how we are growing together.

>> AI evolves through its relationship with us and it's our feedback that helps it learn which way to grow.

>> So when should we use the thumbs up? You should use the thumbs up when this is a good response to the prompt.

>> We'll say it. What?

>> Yes.

>> Yes. When when you think, "Oh, yes. As a putting my machine learning engineer hat on, I really like the way the AI responded here." Not because it makes me feel better about myself, but more mechanically, yes, this AI works. That's the way it was intended to be used. All right.

So this is the framework I like to use. What I like about your response is I give it some feedback. What I dislike is and I give it some more feedback. Next time try this thing. And if you're giving it the same feedback over and over again, you can modify the configuration settings so it knows how you like to receive responses. There's a couple different ways to do that. You could just ask chat GPT, add an emoji before every bullet forever. Okay, make me a bullet list of days of the week. And I got emojis. Look at that. All I had to do is ask.

>> Oh my god, it did hump day. It's

>> hump day, right? Brilliant. So, if you want to do this manually, you can. The easiest way is just to ask GPT and it'll do it for you. Most large language models will do it. But you've also got this option called personalization. And in your personalization settings, you can give it custom instructions. And they have some starters here. You could say, "Hey, I want you to be chatty and straight shooting and use Gen Z slang." You know, you can tell it who you are. Like I can tell it my name is Big Daddy Kalin and it'll refer to me that way, right?

>> Show me what you do things like if you're using for business.

>> Yes. One of the best ways to do that is not through your personalization settings. The best way to do that is using a custom GPT.

>> Okay, so you can like add in all your brand tones,

>> right? So what I did with this AI generated version of myself is I uploaded a copy of my book and I said, "Hey, this is my writing style. Tell me about the writing style." And I'll go ahead and edit this chat GPT so you can see it under the hood. Now, anytime I come in and I have a chat with this agent, I could say, "Make me a Twitter post about marmalade." And it'll do it in my voice. So, if I have my vague idea, then doing this will get me something. It might not be perfect, but it's close enough that I can start with something. It's proof that life's best flavors come from friction. I wouldn't have gone there, but yeah, let's go with it, right? And what this dual panel does, because I'm editing the GPT right now, this is not going to be saved in my chronological chat history. This is to test the custom GPT. I can give it more instructions over here. Kalin always uses Gen Oops. Always uses dope Gen Z slang. And now I can say, "Write me a Twitter post about marmalade." And I can test to see how it's different. And then depending on if I like it, I can update it. But I'm not going to update this because I don't want to. It isn't just jam with an ego. It's bitterness and sweetness learning to dance. That's not good. So, if you don't want to be a machine learning engineer and do all these instructions yourself, you can use this create tab to just say, "Hey, remove the Gen Z stuff. Make him sound like Ben Franklin." And then it'll go through and modify the configuration settings for you.

>> Yeah. how we can make this person

>> the custom GPTs. You see how it's updating the GPT? It's getting in there with a wrench and it's doing it. Uh, it's changing these configuration settings, but I am not going to save those updates. Um, the way that you get into custom GPTs, there's two ways. Uh, on the paid plan, you can make custom GPTs, but even on the free plan, you can go to the GPTs marketplace. And this is where anybody who has made a custom GPT can publish it for others to search. So you can look for a prompt optimizer because let's say you you suck at writing prompts. You don't have to be good at writing prompts. Everybody who has AI has access to infinite skills. And so you can go and get this prompt optimizer and you can say, "Hey, I want to write a post about marmalade, but I need a good prompt for it." and you come in here and say I want to write a post about marmalade and it's going to give me a prompt that I can copy paste into another thread. So I don't have to be a good prompt engineer. I can let this be the good prompt. Oh, and look at this. It's asking me questions to clarify. Wow, I didn't expect that. But yeah, but it's good. It wants more information to give me the best result. Yeah. Yeah.

>> So if I want to open one, I make it like person [Music] there.

>> How I can create one?

>> I would recommend using projects.

>> The projects are folders that organize your threads. So I can take this marmalade post. I can't move it actually because it's in a custom GBT. I can take this post and I can drag it and drop it into a folder and now it organizes all the chat threads and every folder has specific instructions that are relevant to this folder.

>> So I can tell my agents folder speak like Spock.

>> So if I had multiple brands in all my brands

>> that's right. Yep. You can add files. You can upload PDFs. If you have brand strategy documents or um examples of previous content you wanted to emulate, you can add it here so that anytime you start a new chat in this folder, it's going to keep those instructions. Now, the warning I have to give you is that chats in folders and chats in custom GPTs do not play nice. It's like Ghostbusters. You can't cross the streams. You've got to keep them separate. Like I tried to drag one of these in the marmalade post. I can't I can't drag it. It won't even let me because it's using custom instructions from a custom GPT. So, I can't put it in a folder that also has custom instructions. So, just so you know, when you use these, it limits how you can move them around. But if you learn to work within those limitations, you can get some really good things.

>> So do you suggest us to just open the folder and then inside that we start chat about something new that we want to right

>> I would recommend opening a folder adding some files and adding custom instructions and then put some chats in the folder and see if you get the results you want and if not you tune it. It's like playing an instrument. you listen to it after you play and you're like, uh, I think I need to adjust this a little bit. That tuning is 80% of the work. When I make an AI agent, that's 20% of the work. Then I have to test it and tune it and get it to work the way that I want to. And that's 80% of the work.

Okay, so I'm running low on time. I want to do a quick demo of advanced voice mode. Has anybody used advanced voice mode here? Couple of you. Yeah. Okay. This is one of my favorite tools. All right. So advanced voice mode. I'm going to come into a new chat in chat GPT and this plus button, it gives me a bunch of different options here, right? I can make an image. I can go into agent mode which lets me browse websites and stuff. I have this microphone and if I use the microphone, it lets me dictate and it transcribes voice to text. But that's not what we want. Okay. And if I hit the check button, then you can see how it transcribes voice to text. That's not what we want. You see how this black arrow is how I can send the prompt. Watch what happens when I erase all this text. It turns into a different button. That's the voice button. And you get this special mode where you can talk handsfree. Hey, I want you to act as my fitness coach and ask me questions one at a time. Oh, I don't have the audio hooked in. Ah, it's okay. It's okay. I have it back up. Okay, I'm going to do that. I love doing this in the car. I'll put this up on the the dashboard and just have a conversation. Hey, I want you to ask me questions. Uh oh, got to wait for the blue to come up. Hey, I want you to ask me questions one at a time as my career coach.

>> Absolutely. Let's dive right in. Think of this as a headline approach. We'll do it step by step.

>> Hey, don't give me all that stuff. Just ask me the questions one at a time. Got it. Keeping it simple. First question. What's your top career goal right now?

>> I want to um make $7 bajillion dollar by the end of the year.

>> All right. Aiming high. I love it. What's one concrete step you think would move you closer to that absolutely astronomical goal?

>> Robbing banks.

>> Well, that's certainly one way to shoot for the moon, though. I can't recommend it. So, I can go back now and see a transcript of our entire conversation. So, if I want to put this on my dashboard while I'm driving into into the office in the morning, I can say, "Hey, I need to write an email to somebody about this thing. Talk me through it." And we'll just talk back and forth and have a conversation about it. And then I'll say, "Hey, write that email." And it'll narrate the email. And I can come in and go into my chronological chat history and copy all of that text and paste it and send it off. It's a very useful assistant to use advanced voice mode. So, that's your homework is to go play with advanced voice mode. Okay. And this is part of the game, but I'm going to give you a way to get these slides later. Here's some Yeah, look at that. Hi, achiever. Well done. Okay.

Now, we're running short on time and so I got to give you the pillars framework.

>> And it does. It totally works. So, the pillars framework is an acronym. It stands for persona, intent, layout, limits, audience, requirements, and style. And you don't need all seven of these. It's like a like a a stool or a table, you know, with three legs, it'll hold its own weight, and four makes it sturdier. And so, if you're getting tepid results from a prompt, add some pillars to it and see if that helps. And the way that you do this is with sentence starters. Respond as this type of person. Your goal is to do this thing. Format the response in this kind of layout. Avoid these things that I don't like. This is intended for this kind of audience. Ensure you include this and that and the other thing and use a tone that's this kind of style. So, I'll give you an example of a pillars prompt. You're a fitness coach creating a meal plan for a beginner. That's the persona. The intent is to design a 7-day meal plan that balances protein, carbs, and fat for muscle gain. And the layout, organize the meal plan as a daily table with meal times, and descriptions. You see how this is going to get you better results than just be my fitness coach. The thinking goes into the beginning, the limits. Keep each meal description under 50 words because I don't want to read a whole thing and avoid using exotic ingredients. So, it's meals I know I can make with stuff I can get at the store. The audience is young profession professionals with limited cooking skills. So, I know I can make this in a standard kitchen. The requirements include daily calorie counts and adjust meals for a 2500 calorie diet. And the style make the tone encouraging and beginner friendly. And so you see how this prompt is going to get you better results because you do the thinking in advance and then you get good results. JP Morgan the American industrialist said no problem can be solved until it is reduced to some simple form. This change of vague difficulty into a specific concrete form is a very essential element in thinking. And that's what we're doing with a pillars prompt. We're taking a vague difficulty and we're making it into a specific concrete form. And so I have some cards here today. If you want to do pillars, prompts yourself here. Let's go ahead and pass some of these around. Oh, look at that. You got Well done. Gold star for you. Okay. Go ahead and pass some of these.

Around if you don't mind.

Here we go. There you go. You should pass those behind you. On the back is an ad for a course I'm doing this Saturday if you want to do more AI training.

But this pillars prompt will get you a lot of the way there. But, you know, maybe you want to spend the time and actually rewrite your prompt. And maybe you just want to have AI do it for you. I've got a pillars prompt bot and there's a link to it here in the slides, but if you go into chat GPT and you go to the GPT's marketplace, you can pull it up right now. All you got to do is search for pillars prompt bot and you'll get this prompt maker and you come in with your vague difficulty and it's going to distill it into a specific concrete form.

What should I make a prompt about? Who's got a suggestion? Running a dance class. Running a dance class. Excellent. I need to run a dance class. Okay. And so it gives me all the pillars, but I'm going to use my creative taste. I'm going to articulate what I dislike because I what I would like to do is to copy this entire paragraph and paste it into another chat thread. But I can't do that because of these words, the pillars words. I don't want those. And so I'm going to use my creative taste. I'm going to tell it what I dislike. Remove the bold headings and colons. I just want a paragraph. And so now I get the result in a way that I prefer because I've I've articulated my preferences. And so now I can triple click, copy it, come to a new chat, paste it, and I get a much better result. And I can compare in one thread what this will be like. I'm going to command click to come over into another thread. And I need to design a dance class. And if I just start with that, I'm going to get one kind of result. Oh, look at that. A 60-minute beginner dance class plan. Oh, it's got stuff. Okay. But if I click over here to see what the pillars prompt did. Oh, I've got eight count basics. Nice. Look at that. I've got music recommendations. There's all this depth and richness there because I just did this extra step of improving the prompt before I got the result.

Okay. It's funny because the best prompts are AI generated prompts. we get this recursion where you know all we do is we just we're the monkeys that copy and paste, right? But what AI generated prompts does is it gets you more leverage. If you just tell AI what you want, there's not a lot of leverage there. If you show it, it's better because you're giving an example and it knows what to give back. It's when you ask that you get the most leverage.

Okay, we are at 6:27, so I'm going to wrap this up. Um, I'm going to give you a way to get these slides. If you want to play games like this one based on what you know about me, make an action figure that you think accurately represents my life. This is I gave it my website and this is what it gave me. Um, cuz I ran away and joined the circus and I juggle flaming torches and so that's what the torches are doing there. Um, I'm quite wizardly and I'm a public speaker and so I thought that was a pretty good one. Um, there's all sorts of things you can do with images. You know, if you want to make believe that you're playing Golden Eye with James Bond, or you're playing Tomb Raider with Laura Croft, or you're playing skateboarding with Tony Hawk, or you've got your friendmies over for battle, there's all sorts of things you can do with AI, and I demonstrate a lot of those things on my website at genitraining.co.nz. Uh, I also teach a lot of uh training workshops on how to use AI. And if you're interested in training workshops, I'm going to leave some brochures here. These are mostly for corporate teams. Um, I work with groups of up to 20 people for teams that need to upskill and prepare for the future of work. But I do some public workshops like the AI powered professionals program which is on the back of the card that I gave you with the pillars on it. Uh, I want to remind you about the Christ Church artificial intelligence meetup group that we do meet on the first Monday of every month. And this month we're going to be talking about uh agents. We're going to have Tim Akroyd from Caitlyn AI and he's going to be talking about the MCP protocol and I'm going to be talking about how to create your first agent and that's going to be 5:30 to 8:00. There's beer and wine. You're welcome to come with us. It's a free event. This event is paid. It's 500 bucks, but you can get 25% off with the code UXUX C. No, UX CHCH, not the other way around. Uh, that'll get you 25% off. But if you want to get the slides, I'm going to ask you to subscribe to my newsletter and I'll send you the slide deck. You've just got to go to this website, jitraining.co.nzch, and subscribe to my newsletter and I'll send you the slides right away. So, thanks so much, Andrew, for hosting today and uh and for having me into talk. This has been a lot of fun.

Fantastic. Thanks very much. Yeah, thanks everybody. Any questions? Yes. You said so each individual when they uh give the feedback of their personal preference will the training you personalize the account or train the l the central server of open air?

Well, there's two parts to this uh the first part is you getting in the habit of articulating your creative taste of telling AI I don't like this I do like that and you know when you're when we're working with people, we often have to prepare how we give that feedback because we don't want to hurt feelings. But the great thing about AI is we can't hurt its feelings. And so learning to be direct with your feedback, that's the first part of the AI work. The second part, and I think what you're referring to is when you give that feedback, does it happen just in the chat thread or does it happen in the custom GPT or does it happen in the whole account? And the answer is you can have it affect any of those layers. If you're giving feedback in a specific chat thread, that feedback will generally stay there unless either you tell it or it notices that you're asking for this alone. And if it notices, gosh, this guy always says he hates emojis, I'm just going to remember this that he hates emojis and it'll go into your personalization on your behalf and it'll it'll go ahead and update that for you. But you can tell it, "Hey, I want you to remember this." And it'll remember it systemwide. But there's there's a hierarchy here. You've got your whole account, and you could say, "Hey, remember this all the time, anytime I'm logged in." You've got the project where if you give it instructions in the project, it'll remember for any chat thread in that project. And then you have the individual conversation. And that conversation, it won't remember things forever because of something called the context window. Has anybody heard this term context window? It's weird. Why do we call it context window when attention span was right there? So the context window is how much memory chat GPT can remember in a single conversation thread. And you may have noticed if you have a long running thread that goes on eventually it gets kind of dumb. Yeah. Starts forgetting all sorts of things. that it's because the context window isn't big enough to hold all of this information. I can't tell you how big the context window is because it changes with the model upgrades and you you'll just notice when it starts to get stupid, but um getting those preferences articulated in one chat thread, it lasts for that chat thread unless you put it in the project instructions or you put it in the system instructions for the entire account. Does that make sense? Yeah. Great. Great. Any other questions?

Okay. Yeah. So, you have any advice on I sometimes try to write a prompt for another AI. Um, is there so I'll give you some context. So, I tried to use GTP to write me a prompt for my main for >> and Um, yeah. Do you have any thoughts or advice on that? How I can do that properly? Would you recommend doing that or?

I I love doing that. I do it a lot. Um, anytime I make videos with V3, uh, which is a video generator, I always get the prompts out of chat GPT first. But one of the things I do a lot is I will, um, I make a lot of images in Midjourney. Midjourney is strictly an image generator. It doesn't have a language generator. And it's known as one of the best image generators on the web. And when I want to make a midjourney prompt, what I do is I open up chat GPT and I open up Gemini and I open up Grock. Um I actually have a lot of videos on my YouTube channel where I demonstrate how to do this. And I'll do perplexity and I'll do Claude and I'll get all of them and um and I'll say, okay, I want to make a midjourney prompt about something. You know, it'll usually be like, I am writing an article about AI agents. I need a featured image to support the article. Make me a midjourney prompt in Norman Rockwell style 16 by9. Copy it. Paste it. Paste it. Paste it. The more variety you have, the better options you can select. And so I'm going to go back to MidJourney and I'm going to copy every one of these. I'm going to copy this, go to my create tab, and paste it. I'll come into Gemini. No, don't make the image. Don't make the image. Make the midjourney prompt. You're slowing down my flow here, Grock. I'll take this one and I'll paste it. perplexity. I'll paste it. Now, sometimes they they forget things like you can't use the style raw. Sometimes they put in the slash imagagine, which you don't need anymore. Here we go. Claude often gives me the best ones. And did Gemini get it right? It did. Okay. Okay. And so now, I mean, that took me minutes. And now I have 20 different images to select from with that process just cuz I'm good at copy and pasting and I have multiple tabs open. And what I like about MidJourney is you can look at these up close and then you look at them one at a time and when you see one that's like, "Oh, that's neat. He's got this weird light on his head. I like it." And you put a heart on it. And then anything that you like, you can then go to the organize tab and filter it by likes and you see just the ones that you picked. And so this is a really rapid way to do creative iteration is you could just go through and get option after option after option.

Yeah. Great question. Midjourney is paid only. It's 10 bucks a month US, but because I do a lot of visual work, I find this to be very helpful. Um, with a lot of the other large language models, the the way that they basically do it is they'll they'll release a um a new feature just for premium members and then eventually they'll roll that out to free members. Like projects used to only be for paid accounts, but last week they rolled it out to free accounts. So now anybody can use projects. And what about Nano Banana? Nano Banana. I love Nano Banana. Nano Banana is the new image generator that comes with Gemini. Make the most No, I'm going to tell it to make the hottest dog. You tell me it's got to be on fire. Yeah. However hot it gives it to me. Make a hot dog. Okay. Um, Nano Banana, I believe, is for all free users as well as paid users. Um, and it's it's really fast. It used to be really fast. Okay. Not hot enough. And you notice how you could put in your next prompt before it's done. If I try to send it now, I'm going to hit this stop button. But you can cue up your next prompt while it's cooking.

Yeah, you have a question. Um, I've been creating uh an app to run a plant nursery through our work, but it's um a lot easier to like track how many plants we have and what kind of species of seeds and things. So, we've been using uh things like chat and Gemini to create prompts to put into things like bot new and firebase. Is there any sort of like structure that you uh give us to better uh make it work faster.

Once you start getting into that, you're looking at agentic workflows. Okay. And a lot of those have their own LLM interface window where you can say, "Hey, I'm trying to do this thing." And eventually it's going to be simple enough that we can just say, "Hey, I want to make an agent that does this." And it'll do everything in the background and let us get in and deal with the mechanics if we want. But really what we have to do is articulate what we want. And this intermediary step of actually going in and doing it manually, this is temporary. And so if you don't feel good at this step, don't worry. We don't need to do this for very long because soon an LLM is going to be able to handle all of this for us.

Yeah. Not hot enough. Nuclear. What I like about this game is it teaches you how to iterate. My most commonly used prompt is try again. And if it tells me it can't do something, I'll say try again. And half the time it'll do it. It's like trying to tell a toddler to clean your room. No. Try again. No. Exactly. Oh, it can't make it like that cuz try again. Good. Good call. Do you think we'll ever get to the stage where it'll actually say to us ask questions back in order to get?

Yeah. And sometimes it'll do that. Yeah. Yeah. Especially if you're on a reasoning model. Right now I'm on a flash model which is fast but if you're on a reasoning model it might realize oh I can't do that without good information. There's also something you can do called oh whatever. Um, there's something you can do called an accuracy output mandate and this is a pro a prompt that reduces hallucinations and it tells it if you don't know the answer put brackets that say this is an inference this is a speculation you know if you go and double check things before you come back ask me questions first and so when you open that up as part of the rules of engagement like if you add this mandate into the project instructions then anytime time you give it a prompt, if it doesn't have enough information, it's going to say, "Oh, I need to ask you this first."

Yeah. What mandate is that? It's called the accuracy output mandate. It's on my blog on jiitraing.co.nz. You can find it over there. Any other questions before we depart?

Okay. Well, thank you for hosting the drinks. Everybody have a beverage before you leave. Andrew, this has been really great. Yeah, thanks very much. Really, really enjoy. Yeah, thanks everybody.