Loading summary
Transcript75 lines
- [00:01]
A
Let me ask you something. When your team needs AI guidance, do they come to you? When leadership asks about AI strategy, is your opinion the one that matters? If you hesitated on either one of these questions, you're not alone. The AI revolution is creating a new hierarchy in marketing. Those who master AI are becoming indispensable. Those who don't are becoming replaceable. AI Business World Business positions you on the right side of this divide. Two focus days in Anaheim, California, April 29th and 30th, designed to transform you from quote, the person learning AI unquote into quote, the AI expert everyone depends on, unquote. Melanie Miller told us the AI teaching was mind blowing. You'll master workflows that deliver measurable roi, learn from practitioners already providing results and and build a network of 1000 AI focused professionals. This is more than just learning new tools. It's about professional security, career advancement, becoming the person your organization can't afford to lose. Learn more@AI businessworld.live. get your tickets at a businessworld.live. welcome to the AI Explored podcast, helping.
- [01:24]
B
You put AI to work.
- [01:27]
A
And now, here's your host, Michael Stelzner. Hello, hello, hello. Thank you so much for joining me for the AI Explored podcast brought to you by Social Media Examiner. I'm your host, Michael Stelzner, and this is the podcast for marketers, creators and business owners who want to know how to put AI to work. People say there are magic prompts that are guaranteed to get you better results, but is this really true? In today's episode of the AI Explored Podcast, we'll explore a new way to think about prompting. My special guest is an AI strategist who helps companies companies adopt AI tools. He's the founder of Everyday AI, a media company and consultancy. He's also hosted the Everyday AI podcast and newsletter and his courses Prime Prompt Polish. Jordan Wilson, welcome to the show. How you doing today, Michael?
- [02:13]
B
I'm doing great. Thanks for having me. There's nothing better than having a daily podcast than get to go on another one such as yours, you know, so double duty. I'm excited.
- [02:22]
A
Well, it's awesome to have you. So let's just get started with how in the world did you get into AI? Like, I'm curious about your journey because I know you did something before AI as we all. Dude, let's hear it.
- [02:32]
B
Yeah, I'll try to give this super short version because I know everyone has their own story. So essentially what happened is I had a marketing agency. I remember we had our best month ever at the time, very small team, I think there was like six of us. And this was back in 2020, and this new piece of technology came out that was the precursor to ChatGPT, but it was OpenAI's GPT technology, and it started rolling out two years before ChatGPT to a lot of these products and these pieces of software that marketers and advertisers use. So I remember, you know, kind of huddling the team around this new piece of software the day it came out, and I was like, okay, you know, this seems promising. And after a couple of weeks and a couple of months, and, you know, I have a background in journalism, I was like, wait, when you start to really understand and use this technology, it's just as good as the room full of us human professionals. And even if you take a step back before that, before started that company in the actual business plan, I said, hey, I'm going to start this marketing agency. I'm going to keep it going for maybe five years. All I'm going to do is save up money because I feel that there's some piece of technology coming out that I don't know what it is, but it's going to make a lot of the work that I'm good at kind of obsolete. So I had no clue that would be generative AI in large language models, but I was kind of planning for it the whole time. So that was my kind of line in the sand, so to speak, using the precursor to ChatGPT and seeing just how good it was. And as a former journalist, I looked at that and I'm like, yeah, this might have been the thing that I kind of thought about in my business plan that's going to signal the shift for the rest of my career.
- [04:16]
A
I mean, I don't even remember what the heck the tool was, but I know there was a tool that kept changing names. You know the one I'm talking about?
- [04:21]
B
Oh, yeah, it was Jasper Jarvis.
- [04:23]
A
Yeah, yeah, yeah, that was the one that was doing all the marketing out there and I started using it. And I also have a background as a writer and I'm like, this thing is just okay, you know what I mean? It was kind of frustrating, to be honest with you, at first, but, man, when ChatGPT rolled out, that was an absolute game changer. So tell us a little bit more about, like, what you're doing now.
- [04:39]
B
What I'm doing now is absolutely crazy. All I do all day is use AI, talk about AI, teach others about AI. It is both the most exciting thing and the most maddening Thing. Because even as someone that spends 10 to 12 hours a day trying to explore the edges of the technology, you can't keep up. Right. I feel when I started this like almost three years ago, I could feel that no matter what happened on any given day, I would always be at the forefront. It's not like that anymore. As generative AI has infiltrated every crack, not just of our lives and maybe of our skill sets, but also just of every single piece of tool or enterprise software that we use. There is AI literally everywhere. So that's all I spend my time doing, is trying to be extremely fluent in whatever AI technology that companies are using the most.
- [05:37]
A
And you are literally going live, what, five days a week or something like that, Is that correct?
- [05:41]
B
Monday through Friday, 7:30am, Chicago time. Yeah. It's fun to do it live as well, right? Yeah, sometimes you got to do prerecorded, but doing it live is fun.
- [05:50]
A
So like, how long are you typically going live for? And you just covering the news? Pretty much. And your opinions on it, Is that kind of what you're doing?
- [05:55]
B
Shows usually 30 minutes. Try to keep them short. Ish. Because I know a lot of people have worked it into their morning commute. I've literally had people say, you know, CEOs of, you know, medium sized businesses say, yeah, I had to like push back my meeting because I was really into this episode and you couldn't wrap it up quick enough. I have a talking problem. No, not just the news, right? So yeah, we have kind of our weekly segments of different shows. We do the news on Monday and Wednesday is more hands on, practical learning. But yeah, we do a lot of interviews. You know, just really covering it from, you know, every single angle imaginable.
- [06:29]
A
Love it. Okay, so folks, we have a real treat having Jordan here because he is on the front lines of the chaos, as I call it, every single day. When it comes to prompting, which is a big part of what we do every day with AI, what is the big misconception that everybody seems to get wrong when it comes to prompting, that.
- [06:45]
B
You should prompt because you shouldn't. You know, prompting for the most part is going after an output. Right. A prompt isn't just necessarily, oh, I'm hitting the enter key and you know, I'm sending tokens to a large language model. Right. Prompting is when you want an output. So I'd say the biggest mistakes that people make is when looking at ChatGPT or Claude or Gemini as a smarter version of Google search, because that's the absolute worst way to use it is if you're, you know, just trying to get a quick answer or a quick, better written blog post or something like that, it is the absolute worst way to use it. An analogy that I like to use on how people are just completely misusing large language models. I say it's like you have a Ferrari, but you just got the Ferrari to shield yourself from the rain, like you're using it as an umbrella. So if all you're doing is prompting, going in there and you know, trying to find one prompt, that's like having a sports car and not driving it. But you know, you're like, well, hey, I got this to keep myself safe from the rain, right? I'm dry. It's like, okay, well, yeah, you could get an umbrella for that. So, yeah, if you're using it like a deterministic search engine, go use, you know, Google's AI mode for that. Right. If you want to really use a large language model to grow your company and career, you shouldn't be prompting at all.
- [08:14]
A
Okay, wow. Well, that's a fascinating take. So I guess my real question is, we're going to introduce your model in just a second. You've come up with kind of a new way of thinking about prompting. And before we introduce what it is, what I'd like you to do is explain why it's so important. Said another way, if people pay close attention to what we're going to talk about today, what are the benefits or the upside when they apply and adopt what we're going to discuss today.
- [08:36]
B
Absolutely. So first you have to go through an uncomfortable realization that large language models, when you use them the right way, when you use the frontier models, they're better than us, they're smarter than us. Right. I was an award winning journalist, I was a Pulitzer fellow. And even before the models got good, when I learned to use it the right way, I was like, wait, this thing's a better writer than me. So first, if you really want to get the most out of a model, you have to go through that uncomfortable realization that, yeah, this thing's better than me. Even if you've been doing something for 10. Right. I've been getting paid to write for 25 years. It's a better writer than me if you know how to use it correctly. So number one, you have to go through that uncomfortable realization that these models are better. And most people aren't going to go through that. But I think in, you know, maybe two or three years, it'll be easier to see like, oh, yeah, Maybe back In, you know, 2023, 2024, I should have come to that realization because I think people looking at large language models like we have the answer and we are just directing them. And no, you need to be going through a proper context engineering or prompt engineering process where you are collaborating or being a co worker with a large language model. And you know, that's kind of the difference between traditional artificial intelligence and augmented intelligence, which is I'm working with the model to make the model better in my use case and the model is hopefully working alongside with me to make me a better human. Right. So it is this cyclical feedback loop. If you're using a large language model the right way, working within the context window to really expand what's possible both out of you and out of the model.
- [10:13]
A
Okay, so here's what I heard you say. I heard you say that first of all, we need to acknowledge that these AI models are way smarter than we are. That's the first thing I heard you say. The second thing I heard you say is that the upside or the benefit of embracing them in a proper way, which we're going to explain, what that is in just a few minutes, is that it's going to unlock some things. Like what does it unlock for you? Just like tell us a little bit about some of those benefits you've had as a result of doing this new way.
- [10:42]
B
Yeah, I mean, what it's unlocked for me is to feel like I have a full team supporting me when I don't. A lot of people, you know, when they see the amount of content that we put out every day, they're like, oh, how big is your team? And I'm like, there's like two other people that help a little bit. Right. Normally when you look at what we are able to produce, it's, you know, you'd think it's oh, there's five, six, seven, eight full time people. Not really. Right. So that's to address the second half of your question there. But I do want to rewind just a bit if I can because you know, as you repeated that back to me, Michael, what I said, I'm like, I already know that there's going to be people like doubting that. Right. When you say, oh, large language models smarter than you. Because what happens is you always see these things online, right? And on social media and someone shares like, oh my gosh, this, you know, this ChatGPT or Gemini, it can't add two plus two or it can't count the number of Rs and strawberries. Right? So this Jordan guy, he's a complete idiot if he thinks these large language models are smarter. Well, in most cases, whoever shares that online definitely is not using it. Right. And there is this concept of kind of jagged AI. Right. And what that means is even as large language models are out there quite literally discovering new science, Right. Solving decades old math problems that have plagued mathematicians since the 60s, there's always going to be use cases on the flip side where if you don't know what you're doing or if you intentionally ask the wrong version of a model, a complicated question or a simple question, it's going to get it wrong. Right. So you just have to understand it's the same thing. Right. If we had modern day Einstein here, we could probably ask Einstein some simple questions about something trivial that, that Einstein would get wrong. Right. But if you work with Einstein and if you kind of go in the right area of expertise for Einstein, you're probably always going to get that level of outcomes.
- [12:38]
A
Yeah, it's funny, I was thinking about Einstein when you said Einstein. I was going to say if you ask Einstein to draw a banana from memory, he's probably not going to do a very good job, but if you ask him to do a math calculation, he would crush it. So my next question is, talk to me a little bit about models because I know that when we were prepping for this, you said, right model, we were talking about how it's important for people to understand that they have to have the right model for the right project, I guess, or something like that. Like, what do we need to understand? Do you believe that we're at a point right now where it's, you can just use one system for everything, or is it more intelligent for us to understand how to leverage different systems for different things? What's your thoughts on that?
- [13:15]
B
Yeah, you can. It's one of the most common questions that I get all the time. And even just people asking about, you know, hey, can I prompt model A the same way as model B? And the answer is not to use an SEO answer, but you know, it depends because all models technically they have a lot of the same training data, but they have some different and unique training data as well. The models were also trained by different humans, Right. Going through the process of reinforcement learning with human feedback. And then also all models have different system prompts. So there's, you know, even just on the surface, there's at least three different layers where things can go very differently. So yeah, you could work quote unquote with the same prompt. And it might do very well in Claude Opus 4.5 and it might do absolutely terribly in GPT 5.2. Right. So you not only have, you know, kind of the different labs and the different training that goes into them, different data sets, et cetera, but then even when you're working within the same system, it's extremely important as well. You know, probably the most popular system is OpenAI's Chat GPT and its CEO Sam Altman said this summer that 93% of paid users at the time when they had the choice were not choosing a thinking model. And I'm sure we'll get into this a little bit more later if you want, but what that means is even paid users, when they had the choice, they chose the dumber model, right? It's faster, sure, but it has exponentially worse results. So, you know, OpenAI's kind of solution to this at the time is at the time they had the, you know, the GPT series of models and then they had what was called the O series and these were models that were more thinking models or reasoning models. So they worked a little bit differently under the hood. So they said, okay, we're just going to, you know, with GPT5 and after, we're just going to put them all under the same umbrella. Right? Well, good in theory, but terrible at the same time because you still have, I think the overwhelming majority of people making that mistake that leads to these jagged results is choosing the wrong model. You know, a lot of people don't understand that OpenAI got rid of this model routing thing a month ago that was supposed to, well, automatically decide, hey, if this is a hard query, we're going to send it to a thinking model. So they made a big deal out of this when they launched GPT5 back in the summer and then they quietly got rid of it. And people are wondering why you're not getting good results. Well, the model right now by default in ChatGPT, unless you're selecting it, it's called GPT5 to instant. It's actually by third party benchmarks, like the 24th best model in the world. Whereas if you choose the right one from GPT five, two, it's the first best. So talk about a huge difference. And this isn't just a technical benchmarking thing. This is your outputs, this is the work that you're probably using for some aspect of your business. Right? If I could have the absolutely smartest CPA in the world, or the smartest lawyer, or the smartest Financial analyst, and I actually intentionally chose the one who was not that good. Right. That's a human error. Right. So much of, I think what I see people getting wrong when it comes to AI and they're sharing about and being like, oh, yeah, our jobs are safe. Look at how dumb these models are. And most of the times I'm like, that's a human issue. Right. That's because you didn't take the time to understand how these models work and using the right model or the right mode at the right time for the right problem.
- [16:37]
A
Problem, yeah. It's fascinating because I'm an active Claude and Gemini user. I also have ChatGPT, but Gemini Pro is ridiculously good, especially with images. It's like there's nothing even comes close to it. And a lot of people don't even realize that they're still using ChatGPT. And Claude is really, really good with writing. You know, it's like, as far as I'm concerned, superior. But people wouldn't know that if they didn't go out and experiment. And I think that's why it's really important to make sure that you are experimenting and you probably even have access to some of these tools as a result of where you work.
- [17:08]
B
Work.
- [17:08]
A
So, for example, you might have Microsoft Copilot, or you might have Gemini built in with your system. You know what I mean? And these things are powerful. So we're not going to get into the technical weeds today because really what we want to do is get into the concept that we're here to introduce, which is how to go about using the right model. And you might not know what the right model is at first. So just experiment, I think, is the key here. But where do we begin? With your model? Your framework maybe is a better way of saying it.
- [17:34]
B
Yeah, thanks for teeing that up. So, yeah, we've technically been teaching context engineering before context engineering was a thing. And, you know, I'll gladly kind of share the difference between prompting and prompt engineering and context engineering, but for almost three years now, yeah, we started doing this free prime prompt polish course back in 2023. So I've done it, you know, probably like 220 times live for like 15,000 people. And essentially it breaks down prompt engineering into the correct way to do it. And if you look, you know, and we can unwrap this maybe later, but you know, if you look at research papers, right, which I know might be a little boring, and try, try not to get too technical, but there's this concept of, of shots. So you know, maybe you've, you've heard of, you know, an example of a, of a zero shot prompt, you know, a one shot prompt, few shot prompt, etc. But essentially all the research papers show that no matter what model you use, the more shots that are involved in a prompt engineering techn, the better the results. And to simplify that for the everyday person, all that means is when you're working with a model, right, you can get the Einstein result or you can get the idiot result. And a lot of it depends on the context that you share with the model before and kind of the shotting or you give a model examples of here's the input, here's the output, here's what's good, here's what's bad and why. So it's similar to if you were fine tuning a model on the front end. But essentially all the models work this way that you, the human, the onus is on you if you want something that is exponentially better than what a model would give you without going through a prompt engineering or context engineering process, you have to actually put the work in, right? A lot of people think AI is an easy button, but it is actually a bridge that you yourself, the human, have to build alongside the AI and have to understand the rules of how the different systems work.
- [19:31]
A
Okay, cool. So let's start with the first word in your context engineering thing which is prime. Tell us a little bit about what that is and why that matters.
- [19:39]
B
This is so important, right? So priming is essentially context engineering. So it's the process that you go through with any large language model. And I've kind of set this up over the years to really work well with chat GPT. But just so anyone's wondering, for the most part you can do this with any large language model as long as you understand things like context window. And we can get into that if you want. You know, it should work really fine. But priming is essentially the process that you go through and you're collaborating with a model, but in this you are just sharing information and you are specifically telling it, no, I do not want an output yet. So the simplest way to explain this, it's a mindset shift. It really is. Right. The example I always give is, okay, let's put you in the position, Michael. Right? You're hiring 100 employees for your organization. You have a couple different ways you can do this. You could just hire all 100. They come in the room, you throw down a big training manual and you say, go to work. Your first assignment is Due in an hour. That's the equivalent of a copy and paste prompt. Those 100 humans are going to be a little confused. Some of them may be very capable. You might get some great outputs. You might get probably terrible, but probably just a lot of confusion, a lot of gibberish. The same thing can be said when you go click that new chat button in a large language model. So through priming, you're actually, instead of telling those 100 employees, here's a big copy and paste prompt, give me something. Now you're going through a process. What do you know? Here's some information. Let's go through some back and forth, just like you would with humans. You take them through an onboarding program, technically, some reinforcement learning, right? You ask questions. Yes, that's good. No, that's not quite right. Let's take a look at this document that explains it a little more. So all priming is, and we can break it down even more into what we teach is refineq. But all it is is you're going through a process where you are sharing context about your business, about the output, about whoever the AUD that you're ultimately sharing this for, about your expectations, etc. So it is literally a conversation as if you had a consultant from a big four consulting company working with you on whatever that project is. But you are training them and giving them opportunity to ask questions and iterate along the way.
- [21:50]
A
Love it. I'd love to hear about this refine Q thing.
- [21:52]
B
So yeah, here we go. We're going to get a little bit into the weeds.
- [21:55]
A
You can rapid fire through it if you want.
- [21:56]
B
Yeah, we'll do it. So Refine Q just an acronym for the proper way that we like to to prime. So this is role. You assign ChatGPT a role. You give examples of the inputs and outputs of what that role should be going through. Then fetch and insights, those kind of work together. So most people don't know this. Large language models, a lot of times they have very, very old data that they're trained on. Even if the knowledge cutoff says something in 2025, a lot of models are trained on offline data sets and the data can be very old. So in fetching, in insights, right, we're telling, you know, in this case, chat GPT, go fetch this information sometimes from a website, sometimes from a PDF you may be uploading in there and then you're telling it to grab specific insights. A lot of people make the mistake of when you're trying to share context with a large language model. Oh, I'm just going to upload this 100 page PDF and it's going to know everything. No, that's wrong. When you're sharing context, all those things are. They're a book on the bookshelf. Doesn't mean I know it or you know it. You have to tell ChatGPT, go look at chapter four and then go pull out these things. So that's the fetch and the insights. Then we have narrate. So there we're telling ChatGPT the narrative of whoever our audience is, whether that's yourself, whether you're writing something for someone else on a sales page, and then we're explaining the expectations and then we end it very explicitly with q. We tell ChatGPT, yo, you need to come back through this and ask questions, right? I don't want you, whatever the output is, which you explain in the role and examples, I don't want that yet, right? I don't want the SWOT analysis yet. I don't want the, the KPI dashboard, I don't want the financial breakdown yet. First, Q, ask me every single question you have based on the context that I shared. So that's essentially priming in a nutshell. You go through a couple iterations of that until it seems like, yes, okay, now Chat GPT knows, right? In the same way you would training a group of humans. And you're like, okay, now these humans know.
- [23:51]
A
Every year the gap widens. The marketers who understand AI are pulling ahead, creating better content faster, automating strategic work and delivering results their competitors Simply can't match. 2026 could be your best year ever because you finally embraced AI. Now you're listening to this podcast, but why not join us for two days of systematic mastery at AI Business World, April 29th and 30th in Anaheim, California. Here's what Melanie Miller told us, quote, quote the AI teaching was mind blowing. I'm so far ahead of so many people that claim to do what I do because Michael Stelzner only brings the best to teach, unquote. Here's what sets apart this from random tutorials and scattered learning, which I know we all use. You're going to get complete workflows, not isolated tips, strategic frameworks that connect everything, implementation plans you can execute immediately and 20 sessions dedicated to build comprehensive AI mastery. These are the best of the best that I have recruited. By the end of the event, you'll have all the AI skills that took others years of random implementation to figure out literally at your fingertips. The gap is widening, which Side are you going to be on? Learn more at aibusinessworld Live. Grab your tickets today at aibusinessworld Live. Okay, so what I heard you say is role. First of all, what is the role? Secondly example, third fetch slash insights, or fetch and insights, which was like, go out there and find information or take a look at the insights that I have attached that are not examples of what you want. And by the way, the examples are the examples of the output that you're looking for, presumably. Right? Is that kind of what you mean by examples?
- [25:33]
B
Exactly.
- [25:33]
A
So it's got a model and then narrate, which is really a way of like describing who the audience is. Right. So that's where, like your detail who it is you're targeting, what their pains and all that stuff are. Expectations. I'm assuming that means expectations of what, what the role is. Is that right?
- [25:47]
B
Yeah. So at that point you're wrapping up a lot. So you're kind of explaining the expectations to Chat GBT and you're saying, hey, I gave you all of this information. So it's kind of like a summary and explaining what you are ultimately going to ask Chat GPT to do, but then teeing up, don't do it yet. Q. There's still questions. You have to go through this and ask me questions.
- [26:08]
A
Okay, so I'm curious, are these done one interaction at a time or is this like one humongous file that you upload?
- [26:16]
B
The way I generally teach this is to do it all at the same time. Because as an example, when you give a large language model, all of that context at once. Well, I should preface this by saying you should be using a thinking model when you do this. Right? Because if you do, then it will be able to have all of this context at the same time and then give you good questions. Right. In the same way. Going back to the analogy of training a hundred people, you would probably send your people an email first or a quick one pager. Right on. Here's what to expect during training week. Right. That's essentially what we're doing with this process here of refine. Q It is one big chunk. I literally break it up into seven different paragraphs. So it's easy. One for each letter, hit enter, make sure you're in thinking mode, look at the chain of thought and then move on to the next that.
- [27:05]
A
Now, just out of curiosity, when you ask it, if it has questions, do you have any tips on how to inquire of? Do you say, ask me one question at a time? If you need Any clarification or what are we focusing on on that Q part?
- [27:17]
B
In general, I think the way to get the best out of large language models is don't set too strict of expectations when you're still in the iteration phase. So if you said give me five questions, well, maybe there's eight questions that need to be asked. Right. I would say, at least when you start out, never set a hard expectation and you really need to explore. Right. And there's actually one thing that should help. Maybe I should have referenced this before because maybe you're listening to this and you're like, wait, this seems like a lot of work.
- [27:48]
A
Yeah, I was going to say. Yeah, exactly.
- [27:50]
B
Right. So there's a preface to this. I call it like five. Five, five. Right. But this is essentially helps you think about the future of work. You should really only be going through this process, a proper prompt engineering process. In general, I'd say when you know it's something that you will be able to reuse or scale. Right. So if it's something you do once a week, great, go through this process. Right. But I literally think of it like, if you could hire anyone that you wanted and it was free, would you need to hire five people? Would you need to hire 10 people? Would you need to hire 50 people? Right. So get very specific and think of it like this. You are training an employee who in the future is going to do one very specific skill set. So you might say, oh, okay, well, I just need something to help with copywriting. All right, well, wouldn't it be better if you could have have 50 copywriters that all had their very specific expertise versus one that was more general? Yeah, probably. So you do have the thing about scalability and reusability when you're going through this process. And like I said, if not, there's a good chance you might just want to use something like an AI mode in search anyways. Right. If you're just looking for a quick answer, maybe this isn't it. But if you're looking for something that can help you think better, strategize better, produce better outputs for whatever you're working on, then it's usually worth going through this process.
- [29:05]
A
Yeah. And I'll translate what I think I'm hearing you say. What I think I'm hearing you say is, look, your objective here is to create a custom GPT or cloud project or a gem in Google so that what you end up with is something that is trained and reusable. Is that kind of what I'm hearing?
- [29:19]
B
You say exactly.
- [29:20]
A
Okay, cool. So we've spent a fair amount of time on the first part of the process, which is the priming process, which I've effectively heard you say this is the most important part of the process. Right. Because most people just go, go to probably the next stage that we're going to talk about a little bit here. Right. So the big benefits that I heard you say is like you kind of implied this, but what this is doing is it's taking all the insights that these models have, which is massive and it's kind of narrowing down the scope. Right. And it's highly customizing it for you. I mean, is that fair to say?
- [29:47]
B
100%. And this is why, right. People don't understand just how massive large language models are. Right. They're trillions of parameters. It's essentially ingested all of human information that's ever, ever been published, even not just outside of the Internet, even on offline data sets. So yeah, that's why you really have to do a lot of work. I like to say you need to make the model smaller, smarter and more specific for your use case and ultimately scalable and reusable. So yeah, I think thinking of it as, okay, ultimately I'm going to use the results of this to create a custom GPT or a project. You know, is a good way to think of it. But yeah, really think of it as a skill set that you're going to reuse, not just an output. And I think it's important when you go into this process to think of it that way. Don't go in looking for an output or needing an output because then you're going to rush through it and you're really not going to give that employee the training that they need.
- [30:38]
A
Love it. Okay, so what's the next part of the process once we get through the prime part?
- [30:42]
B
Yeah, you hit it. Priming is definitely the thing that takes a lot of time because that is the context engineering and sharing. Right. But after that you get into prompting. But there is usually one little trick before you get into there. And that's the concept of, we've always taught it as a memory recall. You know, a lot of the companies are starting to work this into their large language models right now. Claude recently did, it's called context compaction or auto compaction. And that's essentially going through because, you know, large language models do have a defined and technically somewhat limited context window. So the more you work with it, which is what you should be doing technically, some of the things that Maybe you share earlier on in the conversation are going to drop out of the context window. And that's why a large language model might seem, seem awesome one day or you know, at 7:00am and then by 5:00pm you're like, this thing stinks. That's why.
- [31:30]
A
Well, and just to clarify, because you keep using the same thread, right? That's what you're really talking about when you say it stinks by 5 o'.
- [31:36]
B
Clock. Yes, but even within that same thread, right? Yeah, it's going to eventually start forgetting things because it doesn't have, yes, it has memory. You can reference past chat history, all those things, but it's not going to remember all the nuances and fine detail of the conversation. So, so before you move on to prompting, you do essentially say something like, hey, go recall everything that's important, kind of giving you the cliff notes of the determinations of your conversation that then kind of get pushed to the bottom of the context window. Before you go into prompting question, do.
- [32:05]
A
You ask it to just simply recall what's important or to present what's important? I'm just curious if there's a distinction on that in your opinion.
- [32:14]
B
The language that I usually use is I say, please recall every single important fact determination that may be helpful as if you were giving instructions of our conversation to a large language model that had no history or understanding of this conversation. Right? And then, you know, you should check through and you're like, oh, it forgot these couple of details and you might be thinking, okay, Jordan, why not just start there, right? You, the human, should just go and start there. Well, you don't know what a large language model needs for your outcome and neither does the large language model. So it's like you almost have to go through, through that time of working together. It's like, why would you ever hire a consultant, right, A big consulting company and spend, you know, six, seven figures if all they're doing is ultimately saying, okay, you know, here, here's we, we did our, you know, post it notes on the board and you know, here's your report. Well, sometimes you got to go through that question and answer. You got to poke holes and defend your position to get to that point where you know what is ultimately important. But yes, you know, you are essentially leaving with a version of kind of custom instructions, but at least a summarization of everything that's been determined to be important, important moving forward.
- [33:20]
A
So what I heard you say is you appended this important information, which is this memory recall concept before you begin the prompt so presumably up to this point we prepared some information, we presented it to the model, we've asked it to ask us clarifying questions, we've given answers to it, and we're ready to begin giving it something to do, presumably. Right, exactly. And before we do that, because there could be a lot of interactions depending on what happens here in this context. So we want to say, hey, hey, do a memory recall of all the critical information that we spoke about that would be very valuable for, effectively for you to do your work, something along those lines. Right. And it's going to give you some sort of a summary right there on this page. You can quickly look at it and make sure it's right or it's wrong. And if it's wrong, you can tell, oh, you're not quite right here. And then we're now ready for what we're calling the prompt. Right.
- [34:09]
B
You nailed it. And guess what? Luckily, because the priming was kind of exhaustive, prompting is easy. Easy. All you're doing there is saying, okay, keep all this in mind and now for the first time, go give me that output. Prompting at that point is easy. All you're doing is pulling from all of the insights in the context that you gave the model and you're finally asking specifically for the output. It's always good to re outline it, but prompting is easy. That's the easiest part. It's the treating the large language model like a thought partner in the beginning and going through that multi step refine queue process that actually makes prompting the easiest and the shortest part.
- [34:46]
A
Okay, I want to ask a clarifying question. Earlier in the interview you mentioned the concept of shots and shots are examples, if I'm not mistaken. How important is it if what you're looking for isn't the written word? You know what I mean? But you're looking for something that instead is just like an advisor or something like that. What kind of shots do we use in that kind of context? Does that make sense what I'm asking?
- [35:09]
B
Absolutely. So maybe I can share a quick example if that's okay.
- [35:12]
A
Yeah, please do. Yeah.
- [35:14]
B
So this is something, it's, it's actually the example example that I give in the course. So you know, even for me, right. I went through this process kind of on a competitive analysis for everyday AI. One of the things, after I asked for the first output, so I went through the prompt and then I looked at the response and then you kind of move into the polishing phase. Right. So kind of teeing up the third one here, if that's okay. So one of the examples, the opportunities that the model presented was, okay, Jordan, Jordan, you should, you know, do a certified course. And I'm like, okay, that's good, right? Good example. I put so much free information out there, I really want to educate. Great example. Chat, gbt. So when it comes to polishing or shots, right, I go through input, output, good, bad, why? So I said, hey, in my input, I gave you all this information. I told you to act as a strategist, create a SWOT report for everyday. Why? One of your examples of an output was you said I should go through a certification. However, that was not the best, right? So input, output, bad, good, bad. Why? So I said bad. And I said, here's why. I said, also throughout the course of this context, I told you I have relationships with people at Google, I have relationships with people at OpenAI, I have relationships with people at Microsoft. And I also told you, well, I was a professor of AI at the poll, Paul. So in this case, right, you said I should do a course. Well, shouldn't I do that course and somehow get certification through a university since I have all those contacts and shouldn't I collaborate with some of these previous guests? Since I told you in my context that I'm only able to accept 2% of guests that pitch, so shouldn't I, you know, bring some of these guests in that can help educate, right? So that's an example. Input, output. It gave me not that good, but I give an example of what would have been a good output in that case. So I'm teaching it, I'm course correcting it. I'm putting the red ink on the paper. I said, hey, input, output, good, bad, why?
- [37:12]
A
Interesting. Okay, so what is the benefit of this? Like, help people understand why this is so critical?
- [37:20]
B
This is the difference between having a kind of untrained intern for you to work along with versus someone that knows everything in your head. This is why a lot of times for me, and this is going to sound bad in almost most cases before I would ever hire a human, I probably have a GPT first that I'm going to consult with or I have a Claude project or a Google Gemini gem, because I've already gone through the process of sharing literally everything I can, right? Things that are going well, things that aren't going well, performance metrics, et cetera. I've gone through and trained these things to really push me and poke holes and prod, but doing it with my business context and in mind. So when you look at it, that way, I think it makes sense to go through this process for what it truly is, which is using the best version of yourself combined with the best information out there available on the Internet that can help you accomplish whatever goal you're trying to accomplish. I don't know if that answered the question or not. It did in a very roundabout way.
- [38:29]
A
I'll add a little bit more insight to this from my experience. I use cloud projects and the good news about cloud projects is they have memory. And when I start seeing a lot of repetitive stuff happening, instead of going back and altering the system instruction, I just say, hey, stop doing this and update your memory. This is bad, you know what I mean? Or replace this with this and update your memory. And that's effectively like having a meeting with one of your staffers saying, look, this has got to change. That's effectively what we're talking about right here. Right? Is it not?
- [38:59]
B
It is. So I will put this preface out there. You never want to be over reliant on certain features that can change under the hood without you knowing about it.
- [39:09]
A
Okay, that's good. Keep going.
- [39:10]
B
Yeah, yeah. Like as an example, OpenAI and ChatGPT, they were the first with memory when their memory first came out. Fantastic. I think it's fair to say that their memory has degraded over the past couple of months. I actually love Claude's memory, but you can't leave the that to chance, right? Especially something that you technically don't have a ton of control over. I know at least if you're working in projects, there is the section in Claude where you can kind of see the memory and you can tweak it a little bit. So in that case it does work similarly to custom instructions anyways or a system prompt. But yeah, for the most part, especially if we're talking like ChatGPT or if you're using Google Gemini on a personal account, you don't want to leave too much up to a certain feature that, that there's not a lot known about it. You know, it's not like there's a Change log on OpenAI's website specifically about the memory or on Claude's website specifically about the memory. So you don't have a ton of fine tuned control over that.
- [40:07]
A
So then what do we do? Let's take ChatGPT as an example. Like we baked up, let's say a custom GPT that's doing a really good job for us, but it seems to be failing in a certain area over and over again and we've given it some feedback how do we effectively retrain it?
- [40:20]
B
Michael? It's a very important call out because most people, people don't truly understand kind of the architecture underneath and the scaffolding like that stuff changes all the time, right. These models are agentic by default now, which is both amazing but also a little hard for humans to harness. The other thing is you look at a model, right? Let's just say GPT5.2 is thinking and you get something working and you're like, great, now I'm good, I'm on autopilot. Well, what a lot of people don't know is that GPT5.2 thinking model is getting updated probably every other week. And that doesn't mean it becomes GPT5.2 with thinking. So things are constantly changing. I think one of the worst things that you can do is especially if you're very reliant on large language models in your day to day work, which I think technically we should be. If you're not properly scoping and testing things and if you're instead on a set and forget mindset, it can be bad, Right, Because a lot of things can change under the hood without you knowing it. So if you're just blindly trusting results from any large language model, if you're not properly scoping and testing them on an ongoing basis, if you're not doing the basics on observability, which you can do fairly simply. Right. Even just looking at the chain of thought inside ChatGPT as an example, if you're not doing those things. Yeah, actually just being overly reliant on, oh, this worked last week or oh, well, you know, the memory is going to catch it, you hope it will, but you really want to have a little bit more control. Like a lot of people have this human in the loop, which I hate, FYI, I like to say expert driven loops, you need to be driving those loops. Human in the loop is usually like, okay, well when something goes wrong, which human is the one that's responsible, you know, and you really have to take an active role in the relationship with a large language model, the mode, the processes that you're using, and not just set it and forget it.
- [42:08]
A
So does this mean you're advocating to basically re enhance the, for lack of better words, instruction set and to keep a copy offline, presumably? Right, so you've got a copy of it and just keep testing it because if you're not aware that it's changing under the hood, it might seem like it was smart and now it's not and you're so Reliant on the thing that you don't even have the original set of instructions or whatever. I mean, that's kind of what I'm hearing you say you should have a copy somewhere, right?
- [42:34]
B
Yeah, absolutely. Yeah, you should. And there's so many aspects that you have to, to keep in mind, right. Are you a small team, are you a large team? You know, do you have, you know, technical staff members? But if you are a larger organization or even, even a medium sized organization, yes, you should be keeping these things offline. But you should have someone that goes in at least once a week, especially if you have. Which happens, right. A lot of people, I'm not afraid to call it out. A lot of people look at AI as a shortcut. And yeah, someone on your team builds a great GPT, you're able to use it and you just kind of take a back seat, kick your feet up and like, all right, it's going to be good. You need to have that's going in there at least once a week, going through your different use cases, scoping everything, looking at the chain of thought because maybe you built a GPT for your company and people are using it heavily. And okay, well one time it's, you know, calling Python to do some data analysis and then the next time it's not right, it's skipping over that process. So maybe then you need to say, okay, we need to update our Instructions because maybe OpenAI made a change to the AgentX scaffolding, maybe it made a change to how the model handles certain queries. But now we need to tell it specifically to use Python to do A, B and C. So when you see models doing good things that lead to good outcomes, even if it's not in your custom instructions or your memory, you need to denote that you know. And yes, keeping offline or separate versions of these is an absolute necessity. But yeah, you need to constantly be looking under the hood and updating what's working because you also like, okay, well what happens if for whatever reason one of these AI companies triples their prices overnight and all of sudden a sudden you're losing money? Right. You have to think modularly when it comes to building in your day to day processes with an AI model.
- [44:14]
A
I'm reading some people's minds that are listening right now and they're like, Jordan, that's complicated. And I'm like, okay, yes, it's complicated, but let's be intellectually honest. The benefits of AI allow us to almost be superhuman in many regards, right? So if you want to retain your enhancement, you have to do maintenance and, and it's just not a set and forget kind of thing. And let's not discount the incredible benefits we've all received thus far. And if we don't have a system like what Jordan is talking about, we are eventually going to be outpaced by someone who's keeping up with just making sure their model is actually doing what it's doing. Even if you had a robot, you have to oil the whatever every once in a while and charge the batteries. And there's some things that we need to do. And I think, Jordan, you've developed a really cool framework that allows people to at least wrap their heads around this in a, in a way that allows the model to kind of keep up and allows us to keep up with the model and it's pretty powerful thing. Thank you so much for sharing your wisdom with us. If people want to check out your show Everyday AI, where do you want to send them? And if they want to maybe work with you, where do you want to send them as well?
- [45:20]
B
Yeah, the best two places would number one, probably be our website and, and that's just your everyday AI Com. And the cool thing is is, you know, we have almost like 700 podcasts, so you can go watch them all on our website or watch them on YouTube on our website. They're all on there, listen to them on our website. So everything's free. It's a free, unbiased university. So go check that out or just, you know, connect with me on, on LinkedIn. Right. We do the live stream on LinkedIn every single day. I'm fairly active there trying to answer people's questions, all of those things. So those two places are probably the best.
- [45:53]
A
Thank you so much, Jordan, for sharing your insights with us today.
- [45:56]
B
Michael, thank you.
- [45:57]
A
Hey, if you missed anything, we took all the notes for you over@socialmediaexaminer.com a92. Be sure to follow the show on your favorite podcasting app. And if you've been a longtime listener, we would love a review on whatever platform you're listening on. And do let your friends know about this show and do check out my other show, the social media marketing podcast. This. This brings us to the end of the AI Explored Podcast. I'm your host, Michael Stelzer and I'll be back with you next week. I hope you make the best out of your day and may AI help you become more successful. The AI Explored Podcast is a production.
- [46:29]
B
Of Social Media Examiner.
- [46:34]
A
What if you could get year round AI training. That's exactly what's waiting for you with our AI Business Society. To learn more, visit socialmediaexaminer.com AI.