Summary
1. Deconstruction (Reverse Engineering)
2. System Prompts
3. Workflow Development
Notes
Transcript
How to harness this stuff, right? So, like, nothing is also real anymore. It's getting, like I said, it's getting weird. That's me, but that's also me, but that's not me, right? This is Clay's motion control. It just takes a video of you and skins it as someone else. - But hopefully it won't be March 18th and 19th for Affiliate Takeover. - In Miami where I'm gonna be speaking to you guys, we're gonna be talking everything AI ad creative from static to video to YouTube.
Because we have AI, we do more, and we'll fill that time, right? And that's also leading to what I'm calling creative fatigue. Now that anyone can replicate anything, right, you'll see a winning concept, and then someone will take it and use it in like 10 minutes because of people. Right, so this is gonna happen. I'm sure you've all seen these worst story ads. It's the only thing I see on my damn Instagram feed every time I open it.
One person did it, now it's really easy to recreate. Think about that with the UGC video, VSLs. Everything's just gonna start to look the same, right? So nothing is yours anymore. Whatever you did, anyone else can take it. They can reverse engineer it and do it in five seconds. So we gotta keep moving, right? So here's what we know. - We need more creative, we need new concepts at a faster pace, at a higher quality.
Where it gets complicated, right? What tools do we use? How do we combine them? How do you build a system? How do you not get frustrated? So we're gonna solve that problem today and we're gonna go through a lot of things. I'm just gonna try to rip guys, so sorry if I'm moving real fast. But the next 25 minutes we're gonna go through basically the three core skills for AI, what you need to just continuously move with this stuff.
Building scalable workflows to generate anything in bulk. The billion dollar brand framework, the stuff that I use on the internal side of these billion dollar brands, and then vibe coding some creative systems for a massive scale. So how can you do this? Let's start here. Freeport AI skills. This can be used anywhere, for any tool, basically in any medium, at any time. I like to call it the modern AI tool set.
Deconstruction. Most important thing in AI. Making something you already have, breaking it into components. So it's essentially reverse engineering. You can reverse engineer an image, you can reverse engineer a VSL, you can reverse engineer an email, right? Two, system prompts. System prompts allows us to build instructions for the LLMs or for the agents. So it's like basically building a custom GPT if you've ever done that.
but it's a way to make things pass throughout a system without having to be manually prompted. And then skill three, workflow dev. How do we get from point A to point B? What tools do we need? How do we sequence them? If you know how to do these three things, you can basically do anything with AI. It doesn't matter what tool, doesn't matter what platform, you can use it. So skill one. deconstruction, let's start here, because again, if you can reverse engineer something, you can basically build it.
So this is a simple sort of application of reverse engineering, but it's a universal process that can be used for anything. Think about reverse engineering an ad, whether that's video or image. You can also do copy and scripts, which is part about how you can reverse engineer VSLs, right? So I might give this image from seed over to Claude and say, build me an ad wireframe for about anything that's related to seed.
Now I have a template, right? AI works really good on templates. So we give that seed image, we get the template. Now that template can be repurposed 1,000 times for any brand, right? It's just taking differentiator one, where the hero product is, what the headline is. You can control this yourself. You can run this with agents. But that's the idea is now that one piece, that template is the scalable.
Right? So if we're going to do this, we're going to start like, let's go take a step back. We're going to deconstruct an image because if you can deconstruct anything, that's how you build Trump, how you build formula. Right? So if you want to deconstruct an image, you got to like know what an image is, what makes up an image, what could you need control? Right? So essentially if I gave you a photo, it's like what makes up this photo?
So These are the core elements. If I was to take a photo, any image, and break it down into non-negotiable components that are all going to be in every photo you've ever seen, right? We have our shot type and our photo type, basically how the photo is captured. The radical difference between a drone shot and a close-up shot that tell a totally different story. Subject and action, who they are, what they do.
Environment, where it takes place. The color scheme, what colors are represented. Obviously, black and white. still a color so like that's representative technical details if you're getting more sophisticated you can use things like different camera settings or different tricks like that composition how it's visually constructed how everything is put together within an image the scale and text how big like a product is or a person is where the text is how big it is things like that lighting there's lighting in every single image if there wasn't lighting you just have a black square like textures again this
A little bit more of a differentiator, it'll help your images pop a little more, it gives the surface depth and feel. And then details and modifiers are things you can always add. So we can take all these things and we turn them into what I call visual building blocks. Each one of these core components is what makes up an image. So when we know that and we're trying to control how to create an image, we can then prompt for these things.
We'll take those visual building blocks and we'll turn them into like a formula, like a little map lives for AI to fill in the blank. So, you know, again, we just layered it with shot type, subject action, composition, things like that. Now, if I'm going to read this prompt, this is just like a random set of words saying long shot Kodak Lancer rally car tearing through Kangoo's rainy streets at night, disposable camera, slight exposure.
boards, scooters, and pedestrians all balanced left, right, and right. It's like a random set of words. It wouldn't mean anything. But when I run it through an image generator, bang, we get exactly what we asked for, right? And you see what it's all represented here. You have the Kangoo streets in Bali. You have night rain. You have disposable camera look, gifted that little rawness quality. We have pretty high contrast, meaning the difference between light and dark in a photo.
We have our car, the Kodak Lancia. It's the long shot, meaning it's a little bit pushed back in the frame. It's not exactly like full focus. Car splash, obviously we can see that. Left framing means it's off center a little bit. And then we have our motion board scooters and pedestrians and the outlet. So because I'm controlling every little bit of the non-negotiables of what's going to be in an image, I can control almost every pixel.
It's only going to get more powerful as we keep going. That was cool. That was just creating images from scratch. That works for everything. Now let's talk about editing. This is probably the biggest update in the industry in the last two years, was having these image editing models. They're called in-context models. Basically, you give it an image, give it a prompt, and it'll do something with that image.
They're super intuitive. There's infinite possibilities. It's actually pretty easy. You don't have to go over it. The models, I'm sure everyone here has been using Nano Banana in some way, shape, or form. It's super awesome. It's my favorite tool. But there's some other ones that are coming up. See Dream 5. GPT image is good for what you need it for. is another one, a Chinese model, a little cheaper, especially if you're running API stuff, you can get the job done.
But there's going to be more to come. The industry loves this stuff. They've seen the response. They've seen how much money they make. They're going to keep iterating them. So simple structure. Let's just say a new tool comes out, like Nano Banana. How do you figure out how to prompt this thing? It's like a different infrastructure than something like Midger. We have to just These are deconstructive brains, again, and it's called an edit model.
So I have to think about, like, what is an edit? Edit is basically keep this, change that. That's really it, right? So if I have this image of the athletic means package, I just want to keep that. What do I want to do? Change the background. That's really it. Simple stuff. So when we start to build, like, structure or infrastructure around this, we can start to build prompts where it's like, okay, keep colon edit.
What you want to keep, change, Poland. What you want to change. You can get way more robust with these prompts. This is a really simple one, it works all the time. So, like the example might here be like, keep this exact product visible and preserve the fine details. That's it. Change. Cyclist, mid-training, athletic, you know, AG1 bottle, thrust of the camera, wide angle, vertical prop, well I've wrote that.
Right? So it's just, I'm just giving it an idea and it's just going as right. Now, When we do that, you see, we have our bottle. We kept the bottle, kept the branding, kept how much liquid was in it, and then we just put it somewhere else. We're editing it. It's not rocket science, right? So really, the thing here that I want to, I wish I could do about like a five-hour presentation on nano banana because there's so many things you can do with it.
It's super powerful. It's really infinite in usage and we're not tapping into everything that it can do because most use it wrong. They just think of it like mid-journey or any other generator, right? So the real thing to know about it, and I'm just going to put this up as a pretty simple sort of like framework for it. It's an LLM and an image generator in one thing. Like it doesn't, you don't just type in like do all this stuff, enter, and it's going to create a picture off that.
It's going to go do some research first. So it's going to go multi-step generation, it's gonna like reason with itself and then it's also able to handle like really complex instructions. Some of my prompts are 1700 lines of code asking you to go through different checks and balances to get the right image, right? So something like this with the engine over there. What I said to it was make performance modifications to this engine and label your work.
It just came back and did it and labeled it with brands and stuff. I'm like, I didn't say to do any of that. I didn't put that in the prompt. So what's it doing? It's going out to the internet, finding information, bringing it back in. So you think about that when you're prompting, you don't have to have like this 17,000 word prompt. It can just be like, go to, you know, X website, find the features and benefits, plug it in, structure it in a way that looks like, you know, Nike's advertising, blah, blah, blah.
It's going to go do research, come back and fill it out. So that's just a little tip. There's a couple simple hacks here for nano banana before we move on to more complex stuff. But one of the things that, I don't know if this pisses anyone else off, whenever I see UGC style images or videos and everything looks so fake, like so fake to the point where it's like, how does anyone believe this? Sometimes when you're trying to prompt just straight up into nano or any of these other platforms,
Like, give me a UGC image. It's going to come out looking flat. Everyone's going to be centered and perfect and there's like nothing that says this is real. So simple hack is I just use stuff from my iPhone reel. Just plug it in as that's the image and just say swap the person. Make them look different. It'll keep that whole iPhone aesthetic. Keep the film grain, the colors, the lighting. You don't have to control any of it.
Right? That's one simple way to do it. Not one picture can be used 5,000 times. Or if you have other UGC that you like in perfect settings or things like that, swap out people. You have a whole library at your hands. So the other thing is grids. You can also input one image and get one image out. Or you can input one image and get a grid. So if you just ask for a 3x3 grid or a 2x2 grid, Because Nano has that reasoning capability, you can say, all right, give me a storyboard of this girl in her morning routine.
It's just going to go and create the storyboard of the routine, but you get that all in one output. Same thing with give me four variations to this app. If you're generating a 4K, those are all still usable assets. So there's a lot of different ways to mold these tools to what you need. If you're also running APIs-- You know, getting four images for the cost of one is pretty awesome, especially if you run it in scale.
So just think about some of the stuff as we move along, right? So we're going to leverage this stuff into video. A lot of video is image to video. We're starting out if you want control, right? It's the same basic principles as image, the same visual stuff. The only thing that's different, if you're going to break it down, is motion and sound. That's the only two things, right? So those are the things that we want to focus on.
Now, If anyone is familiar, I don't know if you've heard about this, some, most of you probably have, a new tool called Seedream coming out, Chinese model on ByteDance. This thing is going to blow everything up. It's going to be really good. These images, I just gave it a few images from an Amazon listing, one prompt, and it's creating a full reel based on that. Just one prompt, cut, cut, cut, do this, shots this, and then we're getting a power washing video or a recipe video with a product in it.
It's your APIs, right? That's all I'm gonna say about that one. You just know, right? We can never get enough of it because it works. It's good and it just seems like the format that won't die, right? So if we're gonna use it, we gotta use it a little bit better. I'm sure you guys have seen a lot of stuff. You can probably pick out what's AI just from looking at it on your feed. The reason we like it is because it's authentic, it's unpolished, it's real.
What is AI? Manufactured, polished, fake. And it looks that way. So again, like I try to think about my customer. I don't want them to think I'm just selling them a product with a fake influencer that doesn't exist. And that kind of like counteracts the whole point of UGC, right? So some of the tips I have, if you're going to do it, I mean, obviously we're all going to do it. But really, like, it's not just a focus on what they look like.
It's a focus on what they're doing. Again, we're breaking down. You're doing a UGC video. What are you doing? How are you acting? How are you moving? Focus on the movement. The little nervous tics. Those little anxious things. Like, I'm up here. I might touch my face. I might look down in the way. You prompt for that. You get it. You ask for it, right? So if there's specific demeanor, if someone's excited, if they want to lean in and tell you something, like, put that in your prompts.
So, again, symbol structure. This is really just... Easy framework all I use is just the scene, what's going on, what are we talking about. Character says this, character sounds like this, character's actions, they're doing this. You don't have to write this all yourself. An LLM with this framework, you just say I want it to look natural like someone's anxious and moving like a real human being and not like a robot, it'll understand.
So again, this is what a prompt light looks like. Right, so the other thing you could say is in the scripted language that you're writing, use different patterns the way that you would talk. Use like, sos, ums, things like that. It breaks it out of that like chapped GPT. It just sounds so fake, right? So, okay, so, throw that in there. Eyebrows lift, blinks naturally, eyes guard down to the left, gathering her thoughts.
Head tilts slightly as she gives herself access. Lips pressed together for me, right? We do that, we get a little bit more dynamic. - Okay, so I just got back from a really big event in Miami and honestly, I'm still processing the level of people there, like serious numbers. I met this guy doing eight figures in E-Com and he just sat down next to me at lunch. And then, okay, the party. So there's this yacht party, right?
All white. - We just want to give a little bit more life to our robot companions that are trying to mimic us. So it'll be more convincing. Bellies is a skill too, which is when we started to get into more like automation and sort of system building. Now, system prompts, what they are, essentially they guide the output. They're controllable, they're predictable, and they can be used across like any LLF.
So they're like custom GPTs that pass information, and that's really what we want to do. This is just a basic framework for building system problems. This can be scaled down into infinity, but really this is kind of like a core of what you're doing. You want to know who you're acting as. The difference between act as an accountant, act as a creative strategist, right? We want to put it in that mindset.
Input, output. What you're giving it, what you want it to do with it. Rules. The direct instructions to follow. The core focus. Where it might get tricky and what we actually need to focus on. The format. We want to output that structure. Is it coming out as a script? Is it coming out as a prompt? Is it coming out as a list? And then limits, that's probably the most important thing is what not to do.
Don't do this. So this is like an example of a system prompt. We're going to use this later. But essentially we're acting as magazine photographer who's good at creating prompts. I'm going to give you a text prompt and an image of a product. So it would have an image and my idea. What I want you to do is create ten more prompts from this idea and then follow the rules below, basically. If I give it something, just make it sound better and make it optimized for what we're trying to do.
The rules, like follow these exact. And then I put in the structure from before, the key change structure to give it a format or a structure to output in. And then the format again, how we're going to format those ten prompts to come out. And again, with an example prompt that says, basically, this is a good prompt, this works, use this. And then the limits. This is important because we want to have no additional context or commentary or thoughts.
I don't want it to say, hey, dude, I just wrote 10 prompts. I thought this was a really good exercise. We don't want that. We just want prompts. Give me the prompts, raw prompts, so this can go from this system to that system. No change because we're going to run that directly into the image generators that don't produce for us. So the ultimate goal is to have it look something like this. I attach my product.
I give it just some loose idea of what I want to do. The system prompt says, I got you, dude. I know what you're trying to do. I'll write the prompt for you, and then I'll send it to the image generators to make it go happily. That's three images. That could be 500, right? So this is the ultimate goal. It's like we don't have to write prompts anymore. We can have the other systems write prompts for us.
Now we can send a workload out. This is again variable. This will change all the time, but if you sort of know what you're doing, it'll be much more helpful as things evolve. This is the next stage. It's more controllable, it's more customizable, it's more scalable. And really, you don't have to skip in between tabs anymore, use multiple tools, just do it all in one space. Now, we don't have to do this stuff in single anymore.
I don't have to type in, give me this image, enter, there's my image, and I have to take it and go somewhere else with it. Just like, let's get a single thought, let's get like 20 outputs, let's get 40, let's get 60. So we have options or we just want to keep producing mass asset branches. We have a thinking batch. Now, one of the tools that we like to use a lot, it's pretty simple and we can get everyone up and moving on it real quick.
It's called Weeby, it's node based. Workflow builder has access to 200 other tools, one subscription, that's it. You just have access to everything via API. Scalable as hell. What's important, though, when you're building these workflows is understanding the outcome. What are we trying to do? And then understanding what tools we need. And then understanding what sequence we need to put them in to actually make this thing work.
So we work backwards. So going back to our example before, this is one output. We know that worked. That prompt worked. That key change structure worked. All right, let's try to scale that a few different ways and see if that structure works. Once we know that's good, then we build our system prompt and we're off to a race. So let's do an example, we'll run through this real quick. So if we're gonna do workflow again, we just wanna ID the end goal, what tools we need, we can build a system prompt, and we build a workflow.
So here, basically, what I wanna do is give it a product, and I kinda wanted something to look like this, just like an average ad creative output, asset output, but I want it to auto prompt. Good, so now we know sort of how we have to build it. Node selection, basically I'll take an image input for our reference image. We'll have a text node for our user prompt, text node for our system prompt, a prompt generator that's going to be called A-L-L-M, you can run A-L-L-M through it, and then an image creator that'll be the animal banana in this case.
So the system prompt again, what we're doing here is we're taking, take my reference image, take my idea, write a prompt, and there we go. So it's pretty simple. It's the same thing we just talked about. What we do is we drop in the nano banana node at the end, connect the product to it, we connect the prompt to it, and we have ourselves a little engine. This thing is usable, and it's usable over and over and over again, right?
Because all I'm saying is give me, you know, a girl on the bleachers holding the AG1 product out. But my next idea could be like, I just need a studio shot of this. So it's just, all I'm doing is changing the prompt and workflow doesn't have to change. Things are basically built. Now, where it gets cool is that same engine can be used for any prompt or brand. Where I swap out the AG1 and I put C4 in there, same output.
We're only going to add-- change the system prompt, we're going to branch it a little bit so we can get multiple outputs. So the system prompt, instead of saying, I want you to do one prompt, I want you to write 10 prompts. I just want the goal to be a little bit different. That's it. Now, again, we have our reference image. Our idea of what we want, like put this bottle somewhere doing something, and then the system prompt will say, okay, I understand what you're trying to say.
Write 10 prompts in the prompt generator. Then this thing over here is called array. That'll split it into individual prompts so that I can be running the 10 nanobinata nodes at a time. And then we're basically putting in one idea and getting 20 images out. This can be 20, this can be 40, 60, 80, 100. whatever you're thinking that's how far we can go one play the other thing about weedy that's great is it allows you to put this stuff in app mode so once you build that thing you don't have to look at all the spaghetti you can just drop the picture in drop your idea hit enter 20 pictures right so this becomes scalable this becomes i can hand this off to someone else on my team to do it once it's filled once everyone comes so skill four this is where it gets weird this is me
What else, you know, what are the things that are repetitive that I can sort of handle? Now, the other thing we have to do is start optimizing for AI, for it. It just sounds so weird for me to say this stuff. But sometimes it's just easy to ask AI, what do you like? Not how do I like it, what do you like? How is it the easiest way for you to find information, for you to do stuff? It'll tell you, and then it'll structure it for you.
So, again, what the hell does that mean? I might give it a spreadsheet of all this information and be like, this is how I like my spreadsheets, and it's nice and neat, my little columns. It might say, "Dude, I really like a Notion page with subpages and auto-tagging and all these different kinds of things. It's easier for me to look through." All of this stuff now that you're going to see with the LLMs, the agents, it comes down to data prep.
It sounds a little nerdy and technical, but if you give it something that's a little easier for it to read, it will perform much better. So it's weird. That's typically what we give it. Charlie Day over there. A little conspiracy board. Here's six PDFs, 30 images, five scripts. Figure it out, dude. Right? That's what we give it. Instead of just like a nice neat page file that it can just flip through and bang, bang, bang, bang, bang.
But what we're going to do is have it run through deconstruction, so I call it forensics. What happens then is once we have all that data, like if we're taking out 500 images, it's going to strip each one of those images into like 56 to 60 points of data. It's going to take that big data dump and it's going to send it to what we call clustering. So it's going to take it and make little like five to ten recipes
out of that data. Then we go through this one extra step called abstraction where we just rip everything out that's like brand related or color related to making universal like those wireframe templates we saw before. And then it'll run through my IP and framework on the system prompts, how to build them, and then build me system prompts for this stuff so we can plug it into something like Weeby and generate like 100 images in one click.
So visual learners, images go in. forensics, rips out all that data, skid the clustering, analyze the data, makes recipes. The next skill, takes them, turns them, makes them universal. Next one there, build the system prompts, I plug it into Weeby, I just did like, I don't know, four months worth of work in six minutes, and then I'm generating 100 assets within 10 minutes. So that's how it can get big, right?
So, - Of course skills that are running though, these are the skills, right? We have a forensics, clustering, extraction, system prompt, right? That's cool, that's like the core infrastructure of what's working. What's also cool is that forensics can now be broken down and branched into different things. 'Cause the rest of this stuff will work, so instead of images, we can build it for video, or we can build it for copy, or we can do it for anything that has to do with relational post processing or analyzing edits, things like that.
What's really important is not the actual pipeline, it's these individual pieces. The pitfalls, the prompt structure, the nano-banana thresholds, that's the source. That's what we want to take out of that, and we want to make that generic so it's not just For this single workflow, we can use it for everything else. So a better approach to me is building these skills and roles. They're universal, but they're specialized.
They can be reused for different projects. So instead of it just being like, here's my end-to-end workflow. It comes in here. It ends out here. I might say, I have this project. And my little agent will go, all right, let me go find all the tools that I need for this specific thing to then build it. Instead of just having it built, you can just build on the fly. Right. So, Claude, Codex skills, if you're not familiar with them, they're just like a distilled process, like a set of system prompts.
And, like, again, it's just like basically a custom GV3M2. Building skills, relatively easy, especially in Claude Code and Codex. Just create a folder, link it to the chat, and then go into plan mode and start talking. What do you want to do? It'll help you flush out your ideas. Then the key components, this stuff is pretty, looks pretty much like we talked about in the system prompt. What are you giving it?
What's the purpose? What process does it take? What it shouldn't do? What it creates? What correct looks like? And then where to put it? This stuff isn't insane once you have these sort of frameworks. 'Cause then when you start to build individual skills, like the clustering, like the forensics, like whatever, We're just building like a little workforce, a little pool. And then they can be strung together into different agents because now they're all like redeployable.
You're just picking things up that you need and you're just using them as we go. So the orchestrator, this is the only agent that I really use. I mean, I have another one that I use, but this is just me talking to my little agent saying like, hey, here's some UGC from one of our top competitors and I want to build prompts to like replicate it. Okay, got you. Let me go find what I have in my toolbox and I'll pull it together and make an agent for you and then run it.
And then we have our little skills pool that we talked about here. So I might say to it, I want to build a static workflow for this or this brand or this whatever. Then it's going to go, orchestrator is going to go find, okay, I need to go find those images. I need to extract the data. I need to go cluster, abstract, and then build a system project. If I say, let's make a UGC video, the orchestrator is going to be like, all right, I need to go download some videos.
Then it's going to go to the video forensics. Then we're going to go to the rest of the process. If I say, I want to take this video and make images look like it, it's going to go and download the videos. Then it's going to extract the frames and run it through the image pipeline. It's all reusable. And then my favorite agent over there at the end is the guy that looks at the project as a whole at the end and says, did this work?
And if it works, I save that pipeline. That just gets stored in the toolbox for another day. Whenever the same process comes up, it's like run through this one. And it works really well. So that's how you can start to build per scale. And what's next? I have no idea. Every day is something new. But I think a lot of what we're doing here is replication. So think about that. Everything is replicating.
AI is taking everything that we're doing or everything that has been done and then taking data from it and reproducing it. We're not being predictive, we're not looking at what's next, we're not finding the next UGC. If you're building something and you guys are nerding out, then that's a tool, that's a product, but can we find the next UGC as a format without having to just replicate what you see?
We get it on the beginning of the curve and we crush it profit-wise and then we exit it before it's done. So that's one to thank you guys for listening to me ramble on about that crazy stuff. Right here, I've got some resources. I have the Weeby workflow in there for the bulk production. And I built a little prompt for you guys if you can drop in a cloud. Hopefully it works. It'll build that forensics clustering abstraction skill for you.