Hacking N8N Agents for Infinite Ad Creatives

Hey, today I'm going to be building a system that can turn a single click into thousands of highquality ad creatives that reproduce winning static image ads using NADN and OpenAI. I'm going to build the whole system live in front of you. Going to keep in the detours and any stumbling blocks along the way and you guys can catch the template in the description. Let's get into it.

All right, here's a demo of the working system. It is a simple AI agent connected to a backend image spinner flow. The way that it works is we have a source folder in Google Drive. What we do is we upload images that we want our ad system to spin, aka make slight changes to. And in order to do that, all you have to do is go on Google images and just grab high quality ad creatives, take a screenshot, and literally pop them in here. Once you're done with that, we have a destination folder where all of the ads are going to be generated in. Head over to the agent, put in a prompt like in this case, light blue, pastel tones, cute style, Saraf and Sans, Sarah fonts, flat and minimalistic design, include outlines, company name is leftclick, and add the logo plus a cute little mouse pointer in the bottom right hand corner of the designs. CTA is build scalable systems today and then generate two samples per image. What this is doing is now running our image generation or image spinner prompt. The way that this works under the hood is we take the original image, we analyze the image to generate a very detailed description using this node right over here. Once we have that description, we then feed it into AI and generate what are called change requests. And then finally, we feed in those change requests into GPT image 1 alongside the original source image. This allows us to vary essentially an infinite variety of things in the source image and make it extremely high quality while still retaining the core elements that made that ad pop. And then once all is said and done, what we've done is we generated a bunch of basically spun versions of the original source ad. The value in this is we still get to stay pretty close to a template that has just clearly and obviously won, right? Like there's a reason why this company Spotify, which I know and love, is running an ad like this. Odds are it's successful. They've done some testing on it. So, we can actually take advantage of that testing money and use it to spin a working ad concept or formula and then post it on our own. Yeah. You know, the idea is we don't just do this one shot perfectly. What we do is we generate dozens, hundreds, potentially thousands of ads and then all a creative agency team does is it just goes through this destination Google folder after sending in the request and then odds are one of these ones that we've generated is going to be sufficient if not really good for ads. You can also theoretically just take all these and hook them up directly to some sort of ad system and then just test them against each other. This sort of iteration is soon to be extraordinarily common in the PPC space.

Hey, so this is a live build. I haven't done any of the building yet. All I have is a hope and a prayer and an idea and a rough scope. My anvas is completely blank and I'm just going to walk you guys through my entire thought process from start to finish. Haven't actually built this thing before. The inspiration behind this system is this. I came across a Spotify ad studio ad and I asked myself, hey, could I spin this with a GBT image 1? And then a few days ago when I was designing an AI graphic design agent, I actually used it as a template and then asked AI to modify it a little bit. This is what I ended up getting. So, I consider this a very highquality spin. Obviously, it's different, but you know, I changed the key elements here. Changed the logo, I changed the company name, changed the um, you know, obviously brand colors, I changed some of the text here. And then we turned this from a real image into sort of a cartoon thing. When I saw this, I started thinking immediately as somebody that used to run a PPC agency, well, what's stopping me from not just generating one of these things, but what's stopping me from generating 50 of these things? And then, you know, maybe one of them ends up being that perfect ad combination that I really, really like. Then I thought a little bit more about how, well, you know, when I was doing this conversion over here, I hardcoded it. You know, I said, "Oh, I want a blue thing. I want it to say automations never felt so good. " And so on and so forth. But that sort of ideation is perfect for AI. So what if I just have AI do all that stuff? I could actually take a template or a series of templates and then have AI ideulate a bunch of changes to those templates and then I could have them generate multiple templates. And then after we generate multiple templates, I can have generate multiple ad variants. And then from there, we could just take all those ad variants and test them against each other back when I was doing PPC stuff, which you know, admittedly I was not incredible at, but I was pretty good with stuff like this was part and parcel of our workflow. So my idea is if we do this right, we can save any sort of PPC agency that does this stuff at scale like 80% plus of their creative time, maybe even their copy time too, depending on how good it is. Okay, so this is not like a joke. This is a major system upgrade. I work with creative agencies all the time. This sort of thing like makes a pretty big difference. And over the course of the next however freaking long this video is, I'm going to run through logic that will allow us to build like a productized solution that we can very easily sell to creative agencies or we could use in our own companies with just like an input and a simple output. Okay, so here is the idea. First, we're going to provide a source creative Google Drive folder to the client and the team. The idea is all we do is we just give people, you know, this box and this box is what they use to interact with the system. Okay, so they put a bunch of templates in this box. Templates being like um you know inspiration basically, other sorts of ads that they really like and so on and so forth. And then the box does some super cool magic. And then what comes out is a bunch of results, a bunch of ads, a bunch of spun copy and creative. So the only thing that they see is this box. have to deal with is this Google Drive. So then whatever is in that Google Drive folder, we're going to make the source material for our templating flow. When you ask the model to create something for you, the idea is we take the input and then we spin it 100 times. We then feed in a 100 variations of those spinned prompts into GPT image one which will output each of those spins and the results are then deposited in a result creative Google Drive folder and then you or the team, you know, is like an advertising agency or whatever, you just pick the best five out of the 100. You don't have to do 100, you could do a thousand. Hell, you could do 10,000. It's just about the budget. And then the more spins you have, the more bites of the apple or whatever you allow this model to do, the higher the statistical probability that one of them will be a standout knockout ad. And you know, like an experienced creative person will be able to see this at bird's eye view. So yeah, you know, whether or not you take this and then you give this sort of thing to a person or whether you take this and then you just immediately publish it, like the world really is our oyster at this point. And uh I guess there's some solid implications there as to like where the job of a designer is going in uh 2025, 2026 and beyond. Okay, so here is the rough scope. Here's, you know, above was just the idea. Here's how I'm actually thinking of doing this thing in practice, converting this grand concept into some sort of workable endflow. I'm going to start by creating some Google Drive folders. Then I'm going to grab images in source creative. I'm going to build an AI agent with a tool call that returns 100 variations of a prompt. And then another agent tool call is going to process 100 images with GPT image one, which will then add to result creative. And then I'm just going to notify the team in some way and I'll say, "Hey, your results are ready. Check out this Google Drive. " And then, you know, like if you want to make any changes or whatever, you can. And we're going to keep it simple. This isn't going to be like super intent. I'm not going to deposit it in their project management system or whatever. We're just going to do this like blackbox sort of thing where it's like in a Google Drive. And the only thing you have to do is communicate with that Google Drive because I find that, you know, the audience that I'm targeting here of the creative agency folk. They usually aren't like the most like technically intense. I don't really want to give them too many steps or anything like that. And this is just like the approach that I personally do anytime I'm building something. I care more about the adherence than I do the actual like complexity of the system. All right. So, let's actually go with it. First thing we got to do is we got to create Google Drive folders. There are two Google Drive folders I'm going to do. What did I call that one? Source creative. So, let's go source creative. And then this one over here will be best. That sounds better. So, do I need anything in Nadn for this? No, obviously not. What I'm going to do is I'm going to go over to my Google Drive. And then what I'm going to do is I'll create a folder. This folder is just going to be called oneclick equals 1,000 ad creatives. That sounds sufficiently clickbaity for us. I'm going to give this a click. And then inside of this folder, I'm going to have two. source folder. Then I'm going to have destination folder. And actually, I should just because Google Drive likes to organize these things like alphabetically. I'm going to go one source folder to destination folder. Okay. And now once I'm in here, what I can do is I can open this ID in my URL over here. Just copy this. It'll be a lot easier for us to find it later. I'm going to call this uh let's see, oneclick equals 1,000 ad creatives. I'm going to tag this as NADN course just because I'm compiling a big course that I'll probably recut the stuff into later. And now what I want to do is I just want to verify can I get the data to this folder. Okay, thing in the folder? So search files and folders. That's what I'm going to do. That's how we're going to start this flow. What you need to do in order to make this work is you need to connect your Google Drive account. In order to do this, head over to Open Docs and then there are a few steps you have to follow. You have to start by creating a Google Cloud Console project. You have to enable your Google Drive API. You have to configure your OOTH consent screen. Create your Google OOTH client credentials. Finish your NAD credentials. I know this sounds pretty scary and intimidating. It's not. Configuring your OOTH consent screen is literally you just copying this right over here to Google Drive. It even tells you as much and then's done a great job of like moving towards a simplified version of connections. And then once you're done, you just click sign in with Google. It'll then open a tab and then you just click the account that you want to sign in with and then you're good to go. So, um that's what I've done. Uh now that I have done this, I'm going to close this. This Google Drive is now connected. And then uh what I'm going to do is I'm just going to find files or folders inside of the Google Drive that I just showed you guys. Okay. And in order to do that, we just go down to folder. We can actually just manually hardcode it. We don't actually need the ID. We can select it using an ID if we want, but we don't have to. What was that one called back there? Number one dot source folder. So now we should find the source folder. There we go. There it is. And now what are we doing? We're just looking for files inside of source folder realistically. So, in order for this to actually find something, like I'm going to click execute. It's going to find nothing. But in order for it to find something, we actually need to add something in here. So, I don't know. Why don't I just find some ad creative example? Just going to go back to images and then let's use Facebook ad creatives cuz I just want to find an example of like a Facebook ad that we could spin really quickly. Uh, this one looks pretty good. You see where it says change your practice management software by zero. Yeah, these guys are awesome. Love this ad. This looks fantastic. I'm just going to go over here, copy this. I'm going to paste this in here. Okay, notice how this image doesn't have to be super high quality. I don't believe it needs to be because I did some preliminary testing as I showed you and uh it works pretty well. So, okay, now that this is in here, we should be able to run this. And when we run this, we should be able to get a file. And we did, right? We got a file with the ID. Now, this is I consider really important. We need a super simple flow where a graphic designer, some creative director can spot an ad they like and just dump a bunch of stuff into this Google Drive. How simple of a workflow is that, right? Insane. So, okay, cool. So now that we've verified that we can actually get the data in here. Our first step here is actually done. I always like to keep track where I'm at. So we just created a Google Drive folder. Now we can get the images in the source creative folder. We finished that as well. The next step is we need to build an agent with a tool call that returns 100 prompt variations. This may sound like magic, but what am I trying to do here? Okay. Well, we have an image. Now that we have an image, what we need to do is we need to get a description of the image because what we need to do is we need to flip that description or spin it a million times uh with AI. and then have AI alter that and then feed in all those alterations into a prompt which is ultimately going to generate us an image. If that all sounds like witchcraft to you, don't sweat it. I'll show you guys what I mean. Okay, very first thing we have to do after this is we have to download the file. The reason why we have to do that is because this file is currently not like accessible to us. What this file is this is just an ID. We're searching for it. So, I'm going to click download file and what I'm going to do is I'm going to download the file and I'm going to get it inside of NAN. So, I'm going to go buy ID. I'm just going to feed in this ID. What we have here is we've now mapped this. If I click execute step, we should get the actual image inside of naden. So click view here. And what we've done is we've taken this from our Google Drive and we've actually like funneled it directly into this canvas. If you guys are self-hosting, it is now on your server and so on and so forth. So once we're here, I can now pin all of these outputs. Now that they're pinned, what I'm going to do is I'm going to go into OpenAI and I'm going to get an analyze an image right over here. In order to connect to OpenAI, it's the same sort of deal. Um, this is actually a much easier connection than anything else. What you have to do is you have to create your open AI account. You have to add about $5 in credits at least as of the time of this recording. And then what you do is you get your API key, which you can get if you just click on this page right over here. They make it really nice and easy. After that, paste it in, click connect, and then you are connected. So anyway, there are variety of different things we could ask AI to tell us about this image. What I'm going to do is I'm just going to select GPT40, and I'm just going to start with what's in this image. Now, because this is a binary file, we have to go um from image URLs to binary file. And then the input data field name is going to be data. And basically, it's just always whatever this is. Okay, this is a really weird convention um but just always pick whatever the name of that little um orange section snippet is. So, what happens if I click this? Well, this is now being sent over to AI and AI is going to process it. And uh it doesn't look like this is working right now because it's saying that we didn't find it. I think this is because you're not actually allowed to pin data outputs. I know it's kind of annoying, but maybe we'll pin these two and then when it downloads it, we won't we'll have the unpinned version and we should make it work. I don't know. I'm just going to click execute. Let's see what happens. So, that worked. Okay. The fact that it's taken some time means it it's probably working. We're going to see. Okay. The image shows a person sitting with a laptop smiling. There's a paper with text and a downward arrow icon overlaid on the image. Below there's text that says, "Change your Pascus management software. The background is blue. " This is pretty good, but it's not as good as it could be if you think about it, right? Like, what is this telling us? I think that personally there's more information on here than just that. You know, the backdrop of the woman, for instance, the woman's hair, you know, what the woman is wearing. Obviously, like this little background here as well looks to be like it's some sort of, I don't know, accounting stuff practice management software. This is good, but I'm not seeing a dark sort of overlay behind it. I think we could probably upgrade this. I'm a big fan of being hyper specific. So, let's just see if we can get this even more specific. And the more specificity we get, probably the higher quality our spinning is going to be. So, what I'm going to say here is what's in this image, describe it extremely comprehensively. Leave nothing out. Okay. Now, I'm going to run this flow again, and we're going to see if this prompt is maybe a little bit different. And if this is sufficient, then it's sufficient. We'll just move on. The image features a promotional graphic with a bright blue background. In the center, there's a layered design featuring a photograph of a person sitting with a laptop. She's wearing a light colored blazer and jeans, smiling at the background of lush green foliage. Overlapping the photograph is a document on paper with text and a grid-like table partially visible. Okay, great. So, now we're freaking talking. This is hyper specific. This describes it as good, if not better, than I would. This is great. This is exactly what we wanted. Now that we have this sort of description, we can actually just pin this output because that description is what we're going to use next. Okay, so if you think about it, there's now a million other things that we could do here. Like, I could use another OpenAI node and then have that generate a 100 variations of this. And I think that's probably what I'm going to do. And then after that, what I could do is I could feed that into an agent and have the agent maybe select the template or something like that. Like we'll see. But um okay, what I'm going to do next is let's go back to OpenAI now. And where it says message a model, what I'm going to do is we'll connect with the same credential resource message and model. Model I'm going to use for this is GBT4. 1. The reason why is because as of the time of this recording, this is the largest context window model that I could do. And in our case, we're going to need a very long context window. Also, this one's pretty high quality and it's not super expensive. It's not like GBT4. 5 or anything. So this is sufficient for me. First thing I'm going to do is I'll change ro to system. And then for prompt I'll say you're a helpful intelligent prompt rewriting assistant. You help rewrite prompts. And now what I'm going to do is I'm going to add a user prompt which is where I actually give it the instructions. This over here is just how it is defined, how it identifies. I guess this over here is like what it does and what I want it to do. And the first thing I would do is I always provide some context. So I've generated an image description. I want you to rewrite it in let's just do 25. Well, I guess for test purposes, let's just do five different ways. No, let's do 20. Uh, yeah, let's do five different ways for now. And we can scale this up later. I want you to rewrite it in five different ways given the following input. And then here, I'm going to have the input from the user that requests a specific thing. So, back in the example that I gave you guys, that would be for instance blue background leftclick add studio mouse icon button text that says this, right? Like this is what I'm going to feed in some highle instructions and I'm going to have it generate me a description of all the things that I want. So, I've generated an image description. I want you to take that image description and then give me five variants of it based on the following input. And I'll say input uh and then in this case, let's see like how do we want to modify this test, right? Like we should probably have some modifications. Let's say we want it bright orange. We want a cartoonish character in the middle. We want text saying instead of change your practice management software, maybe we want like upgrade your systems today, you know, and then like a logo of a company, leftclick logo, bottom right hand corner, which is just a mouse icon. Let's say a mouse pointer icon. Okay, I don't know if this is going to work. We're going to give it a try. Okay, what I'm going to do here is I'm going to pretend that this is my like dynamic input. We're just going to feed it in a static input for now. Return your output in the following format. And we're going to do JSON format. What I'm going to do is let me just open this up and then zoom way in on this. We're going to say variance. And now we're going to have an array. And then first image search modified image description, second modified image description, and then we'll go etc. And it should pick up on that. Okay, so we should have an object now with a key called variance with a value that is an array that says first modified image description, second modified image description, etc. Then rules. Your task is to generate new descriptions that we will later feed into an image generation model. New descriptions based on the changes that we will later feed into the image generation model. But do not change anything that is not explicitly specified. Only change that which is requested. Do not modify the image prompt. Maybe I'll just leave that out for now. We'll just give this a test. Yeah, I think that's probably what we'll probably do. I'll click output content as JSON and then under temperature, we'll go 0. 7. Now, what I'm going to do now is I just want to click execute step. And I just want to see can we generate five variants for now that we can then later feed into AI alongside some sort of template. Okay. A bright orange background with a cartoonish character in the center featuring the text upgrade systems today. Above the character, the leftclick logo depicted as a mouse pointer icon is placed at the bottom right hand corner. On a vibrant orange background, a playful cartoon character stands in the middle of the words upgrade your systems today. Okay, this is kind of cool. But what I wanted is I wanted variance. I didn't just want like the text to be slightly varied. What I wanted were actual variants. So, so take the original input, modify the original description, and then generate four other variants with slight modifications to things like color, style, copy, etc. Do not change company names. Do not massively modify. Massively is a relative term. We should avoid relative terms. Do not make large changes, only small ones. And then after this, what I could do is I actually give it some examples of what I want. Um, if I just give it one example of what I want, then it'll do a lot better than me just asking it. Hey, systems now system I upgrade system. Okay, we should actually have it with modifications. Things like color, style, copy, etc. do not Okay, instead of slight modifications, I'll go modifications should now modify the colors and stuff. Okay, upgrade your systems today. Upgrade your digital systems. Enhance your systems today. Cool. All right, that's reasonable. It's not a big deal. Okay, we're now going to pin this output. And now, what do we do from here? Well, we can actually get into like the image generation side of things. Um the way to do image generation there are a couple different ways. What you can do is you can just copy this. Okay. Then you can paste it in and then under resource you change this to image. And now you can actually generate an image. You do that with GPT image one. The issue is the image that we are going to generate with this is not going to be very good. Okay. And I'm going to show you guys what I mean here. Uh let me just set a hard-coded limit. Let me just unpin this. Sorry. Let me just set an example of what I mean here. You know, our original image that we fed in just to be clear was this, right? And we went and we described it. Now we're trying to regenerate it just with GPT image. But as you can see here, this bears very little resemblance to the source image. Now, the reason why you can't generate images like I'm talking about with this method is simply because the model is way too flexible. So we have to do is we have to constrain the hell out of this flexibility by instead of asking it to do it just with text. We actually have to combine text and a template of the image. Now image might sound like witchcraft to you guys but you can do this. You can actually feed in a template. The trick is you don't say it's a template. What you say is hey this is a source image. I want you to make an edit cuz OpenAI's GBT image 1 currently has an edit functionality. So what we can do is we can feed in a source image. We can say I want you to slightly edit this according to this. and then we'll get the exact same output as if it were using a template or scaffolding or something like that. So yeah, we can't just generate the image. What we have to do is we have to modify it. Unfortunately, just based off my experience with the style guide generator, there is no way to modify it just directly built in. You do have to do a custom API call, which is kind of annoying. So what I'm going to do is I'm going to set up a custom API call. Now I'm going to be kind of lazy. I'm just going to feed in the API call from my previous one over here, but I'll show you guys how all that stuff works. No problem. We're going to generate the image using GPT image one. Okay. So, what is this? This is an HTTP request module. The method is post. The URL is api. opena. com/v1/im images/edits. The authentication is a predefined credential type. If you've never used this before, what that means is you click on it, you scroll down under credential type, you just find the thing you want, open AAI, and then even though you don't actually have any sort of built-in endpoint, this can reuse the credentials that you've created for NAND built-ins, which is super convenient. And then you just select your key here. And I have a bunch of different ones. I'm just going to roll with this. Now, what you do next is you click send body form data. And there are three parameters. The first is form data model, and this is GPD-image- one. The second is nadn binary file and then it's image and then input data. Field name is data. Lastly, we have parameter type which is form data and then prompt. Down over here is where you actually have the prompt. So your task is to generate an image. Here is your prompt rules. Stick as closely as possible to the outlines of the source document. Then change text color, shapes, and styles only if explicitly specified in the prompt. Something is not specified, do not change it. Okay, so this is kind of cool. This is kind of neat. This is one of the ways that we're going to follow that source document. And let me show you guys how much better that is than what we're currently doing. Now, one minor issue here is the way that you feed in the data, I'm realizing, is using binary. And I do not know how to feed data from an earlier binary output into a later binary output. That's kind of annoying. Uh I was struggling with this in my last build, so I might as well figure it out now. So, let's see. Feed NAN data binary to later node. Let's just start with that. Okay, this one is saying, are we able to access binary data from a node way back in a JS expression of a later node? And then he says, nope, that's currently not possible. That's the founder of NAN. That's unfortunate. I do not like that. Um, can bring the binary data in again with a merge node. Okay, let me see here. So, oops, wrong one. So, what I could hypothetically do is I could have a merge node somewhere around here, I guess. And then that merge node would take as input this uh could feed it in over here. Yeah, I mean maybe. Just seems kind of unnecessary though, honestly. Like input two and this would be input one. And then we feed this in here. I do not know. Let me find out. We're just going to rearrange this with option shift t. And then uh just because we are now doing this, we have to deselect, deselect I should say. And then for now, I'm just going to have both of these feed into the merge node. Um it looks like we're currently appending the inputs. That's not really good. Let's just combine them all. Let's not combine. Let's just position. This is input one. This is input two. So we should just like stack the fields on top of each other. Let's just keep everything. Don't really want to combine by matching fields to be honest. I'm just going to keep everything. I don't know if this is going to work, but we'll just execute this. We should now have both run. They should also run. We got the analysis. And then we're going to have the spinning. The spinning will now merge. You must at least define one pair of fields and fields to match on. I mean, that's kind of annoying, but why don't we just match on index? Is there an index back here? I don't think so. There's an ID, but there's no index. So, when we run that again. Okay, cool. So, it actually does look like we have the uh we have the data. We had the JSON here. We also have the binary. So, this should work. This should actually work. Cool. Looks like I figured out how to do that. Fantastic. This input data field name is going to be data. So, oh, hold on a second. Wait a second. I'm seeing two items here. I don't actually want two items. I want one item. You know what I mean? Let's see here. So, this is item one. This is item two. That's not good. You know why that's not good? Because if we get in the habit of doing stuff like this, now we're changing, we're fundamentally changing the output flow here. I guess I could just say execute once and it'll only execute a single time with this, which isn't that bad. Still, I want to be able to reference the other data. So, no, the way that I'm currently doing this is not going to work. Um, we should merge the matching items together. We should enrich all of input one with data from input two added in. Okay, so now I'm just going to pin this. Oh, can I do this? I don't know. Let's just find out. Now the output is just one item, which is unfortunate. We need both items. Try combine. Okay, cool, cool. I think combine works. I don't think we have binary data here though now that this is all just in one item. So I think now unfortunately we have to run this one more time. Okay. And then over here the output of the merge node is a single item with our binary. Now that binary is right over here. So we should actually be able to feed this in now. Uh okay. So what I'm going to do is I'm going to feed in the binary field name which is data. And then in addition to this I also have to feed in another thing. So, here is your prompt. And why don't I just feed in let's just do one of these for now. And just so I can show you guys what this actually looks like when all said and done. Okay. So, let's paste all these together. I actually don't know if this is going to work even. Let's do this one more time then. And then let's run this puppy. Okay. And we do be generating. Here's the thing. The way that this HTTP request works, we do just get a bunch of B base B 64. So, I now have to pin this. Then the last thing we have to do is we just go over here and we go B64 convert JSON to binary data or convert hold on JSON data to binary. So I'll go over here and then I'm going to convert this to file. Okay, then we just feed in the B64. json put output data in this field. Let's execute the step. Okay, now we have the file output here. Let's click view. What do we got? Okay, we have um a very kind of scary looking woman over there. Not going to lie, I think the uh modification there was a little bit rough, but it more or less has everything else that we were asking for. So, just to be clear here, the source data of the image is this. So, we can open this in a new tab. This is now our destination data. Okay, so the source destination source destination. Oh, and I'm actually a little bit zoomed in, so let me zoom out a bit. There you go. So, what do we have? We actually have the vast majority of the information here is the same. And here's the thing that I should mention as well. Like I'm not expecting this to be perfect on the first go, right? Like I'm going to generate tons of these and then I'll let the creative designer pick the best pick whatever the heck they want. And I mean like you know if they have an actress or something that they really want in there, they can actually just like crop her out and then stick her in the new one, right? Like no big deal there. But yeah, you know, there are some minor changes. I don't really like how they left out that little bolded section. I do like how they actually added the logo, which is kind of neat. So yeah, that's pretty fun. I think we're doing reasonably well. Looks like the one thing I needed to modify in order to make it work though was I needed to change the prompt from this general description to make the background bright orange. Swap the text for upgrade your systems today. Add a leftclick logo with a mouse pointer and so on and so forth. Right? So what we need to do is we just need to get the supplied prompts to be in that shape you can think of it as. So, I'm going to head over here to the OpenAI node. And what I'm going to do is I'm just going to give it some examples. And instead of generating a description, what I'm going to do is I want you to take the image description and then give me a simple prompt I can use to tell an AI image editor what changes to make. Okay. So, I'm going to say bright orange background, character in the middle, text upgrader systems today, leftclick logo. Let's say small left click logo. bottom right-hand corner. Okay, so this is the main prompt now. New change suggestions that we will later feed into the image generation model. Take the original input suggest changes to the original descriptions and generate four other variants of modifications like color, stock, copy, etc. Not change company names. Don't make large changes, only small ones. So, you know what? I'm not going to write all this stuff myself. Why? Because I'm super lazy and because hey, if I'm doing AI anyway, I might as well have AI help me. So I will say generate four more variants like this. Adjust color copy etc slightly and then I'm just going to feed my actual request in. Oops. Sorry. No, sorry. I just want you to spin this text a little and then generate four more change requests. Modify color etc. And then I will the color. Copy DC. Only thing to never change is the company name. Cool. I like this. Cool. It's doing a pretty good job. So, what I'm going to do now is I'm going to take these and then I'm going to copy and paste them in. Put all five of these. Replace your changed version of my original with my own into an array inside of an object with a key called variance. Now, we're going to grab that and I'm just going to copy paste that puppy in. No need for a key, just the text in an array. No, no. I mean the description key. That's why you need to be super specific with these things. Whatever. I'm going to copy this. Now we're going to go back here and then I'm going to go I'm going to provide the description as a user prompt and then I'm going to provide uh four outputs as this. So I'll now say variance here going to cut it like that. And then over here what I'm going to do is I'm going to feed in my little base description which is right over here. I'm just going to call this original changes. Now we have like modified change requests, right? Then the input is going not actually to be over here. We can do that input later. Over here we're going to have the actual input. Um actually let me Uh, let me think for a second. Bright orange background character, middle tax upgrading system today. Here we might need two. Okay, actually, let's just feed in that base prompt right over here. And we'll call this input um original description. And then I should actually generate the whole thing. Just feed that in there. Okay, that looks pretty fine. And over here under OpenAI, what I'm going to do is I'll just feed in another input uh which we are going to call original description. And I just now need to get something to put in there. Now, I don't want to use the exact same thing that I was just looking at a moment ago, right? Cuz this doesn't make sense. So, I'm just going to grab a new creative. Snap a photo of your receipt and done. Uh I want something that's just a little bit higher quality here. Here's the Spotify thing that I was telling you guys about. H Let's see. Here is a fantastic business, HubSpot. I love these guys to death. Uh, founder actually used to read my newsletter, believe it or not, Darm Shaw. Really smart guy. Really ahead of the curve with a agents and whatnot, which is pretty sweet. I'm going to add this image in here, and then we're going to do our little spinning with that. Okay, so now we're going to grab that image. We're going to download analyze it. We're going to take feed it in here. We're going to merge it. Then we're going to generate an image using GPT image 1. And then let me think here. So, we're feeding in five variants, right? We're merging it and then we're generating the image. I feel like what we should probably do instead, well, it's tough, right? We're feeding in binary to this thing. What we need is we need some sort of loop where we grab that and then we feed in the binary over and over and over again. But how do we do that? That's the question. This to me seems like a data wrangling issue. Okay, for now we're just going to generate it once. We'll worry about the complexity later. But yeah, like what I want is I want this to spit out 50 or 100 and then I just want it to like vary over and over and over again. Just do this. So, we're going to need a loop in here for sure, no matter what, right? Okay, cool. So, we've merged the data. We're now generating the images in GPT. Very cool. I guess another way I could just get around the merge. Like, the reason I don't like the merge is it's I'm just probably going to forget about it. Like, from a maintainability perspective, I could just have it download the same file again. Kind of be annoying to download a file twice. You know, like developers and stuff like that would definitely be like, "Oh, Nick, what the hell are you doing? You no idea how to build an NAN flow if you're doing that. " But like sometimes I'm willing to make these really simple ass trade-offs just because it's simpler for me. like, you know, oh, I just download the same file again. Well, that's easy. I'm even going to have access to the same ID. And then, if you think about it, I'm just going to download it before I do every image generation. And it's um, you know, it's like a linear thing. Like in computer science, they talk um big O. Uh, so, you know, like O of N and stuff like that. This is just like O of 2N, which is O of N. So, it's not that big of a deal. Anyway, let's take a look at the image. Upgrade your systems. Today, we got the exact same vibe going on that we had with the original. It's just now we got that bright orange background and it says left click. I'm not liking how it's just doing bright orange again though to be honest. I think the original was bright orange. Right. So But you know I'm seeing here that we lost some details. So I think if we just generated this 20 times I think a lot of these problems would have been solved. It even like confabulated a you know a forearm for the guy which is kind of neat. But I think if we just did 20 of these it would just like you know one of them would be really standout. So I'm kind of inclined to just run this 20 times. Um the question is now we need to generate an object which has you know we need to have this with the five variants and then for each of these variants we need to feed it in. You know what like I'm I just want to get this done and then I'll worry about optimizing it later as I do most things. So I'm just going to grab this. I'm going to grab this download file. Feed this in right over here. And right over here, what I'm actually going to do is I'm going to split out loop over items split in batches. Delete this. For each of these, I will download the same file. Then using that same file, I will generate a variant. Convert this to a file. Once I'm done converting this to a file, I need to upload this somewhere. So what I'm going to do is I'm going to go to drive, not pipe drive, Google Drive. Going to now upload the file. loop this back around. Where am I going to upload this file? I'm going to go to um the same drive that I had before. What was the name of the drive? To destination, I think, right, there we go. File name. I mean, I don't know. What am I going to name this? Let's just call this like gen. Maybe test will do for now. Okay. No other properties. I'm just going to generate the image as data. That should be output as data if I'm downloading the file again. Oh, I got it. I got it. That's why I hardcoded the prompt. No wonder the image is trash. Okay, cool. Well, now we should be able to split out the items. The item we're going to split out logically is variance. Now, if I just pin this, too. When we do this splitting, we're going to have five items. Each of these items is going to go into the loop over items node. Then we're going to loop. We're going to download the file. What file? JSON. ID. The question is, what is the ID? The ID is going to be the ID from way back over here, which is kind of annoying to access. Do I have access to it over here? It's going to be this Google Drive over here. So, I'm just going to be cheap. I'm going to paste this in delete this again. Why? Because I can actually access this over here, I believe. I might not be able to. We'll find out. So, what am I going to feed in? Google Drive item JSON ID. Yeah, looks like I'm good. And then here, generate the image using GPT1. What am I going to feed in for my prompt? I'll tell you what. Um, loop over items should get message content. So, we should get loop over items, message content. I don't know for sure, but I think we will. So, I'm going to go here. dollar sign loop over items item message. content item JSON dossage content I don't know for sure but we will see then we convert to a file and then Google Drive 03 okay well let's just give this a go and see whatever happens I don't even know first item Google Drive picked it up fine what sort of input did we feed in we fed in this image. What sort of prompt did we feed in? No prompt. Undefined. Okay, that's fine. We're going to fix it up on the second run. Okay, over here, the data that we're going to need is I'm just going to go back a couple nodes till I find the loop of items. This loop over items does indeed have it, but it's in message content variance, which I forgot. So, this looks good to me. We should now actually be able to do the variance. And now that we've done this, one more thing I'm going to add is just a wait. Just going to wait 5 seconds. Not a big deal. And again, I'm just going to spread these out. Now, I'm going to execute the workflow. I will be cognizant of my token costs, hence why I have the 5 seconds. Also, rate limits. And we'll take it from there. Okay, we generated a few. Now, I'm just going to go check out the Google Drive. You can always find Google Drive stuff with the web view link. So, let me just open this up. Upgrade your practice management software. I'm not liking how tall this is. We should probably specify like the height of the images, you know, like you know, don't get me wrong, this is cool, but uh I'm pretty sure the source doc that we fed was a square. So, we don't want this to be a non-s square. So, let's go back to the destination folder. Let's see. So, this is the original. Weird. Doesn't look like it made any changes whatsoever to that one. So, we should check that out. This one definitely looks like it made a change. That change, right? So, I mean, it's making changes, uh, which is cool. And, you know, not all of them are going to be perfect, and that's fine. We sort of expected that. Looks like there was an image on run number four. Why? What was the problem? Oh, never mind. It's not that there was a problem. It's just that uh Well, actually, there is a problem. It says we have four items. Only three went through here. Maybe I'm just crazy right now. Maybe it just hasn't moved the fourth one out yet. I think it should, though. Okay. I really don't know what this issue was with. Was it the third item? Second item. Looks like the second item is still processing. Or is it the fifth item? Hm. Anyway, we got a fourth one here. It is still a little too tall. So, not happy with the height of them, obviously. Uh, so how do we fix the style here? We should be able to fix the style. So, let me just Google this really quickly. Open AI and we'll go this. This will probably get docs for us. When we have the docs, we should be able to take those docs and then find out what exactly is going on here. That's the crate image. They have a size parameter. How about the edit image? Okay, here we go. H. Can we set the size? Yes, we can. Size the generated images must be one of 1024 1024. So, we have to do some sort of 1024x 1024. Um, all right. Let's go back here where it says prompt. We'll now go size. We'll just go 1024* 1024. This should uh solve that specific issue that we were having where it's just all crazy. And then we also just want to avoid them, you know, it having the exact same thing over and over again. So, let's check that out. Yeah. Okay. I mean, you know, like we are getting a ton of different vibes, which is cool, right? I mean, like this one's really cool. This one's actually looks pretty real. No major issues here whatsoever. Um I'm just not a fan of how it's taller, but you know, if this was a square like this, for instance, this would be perfect. This would be a great ad. Simplify your practice workflow. And then like hell, we even got like the Apple uh little Apple MacBook stuff there. And you know what's crazy? It even like almost took the same source person. And here it even This woman is clearly fake and you know invented. But this guy here might be this woman here is clearly inspired by that. So listen, this is good. I like this. Why don't we move this to trash? And why don't we now rerun with the 1024 x 1024. Let me also see how we can improve this. Okay. So they are actually giving very different changes. So I think what we need to do is we need to say first change request, second change request, third change request. You generate new change suggestions that we later feed. Take the original input. Okay. Make these changes minor. We want to stick to the source closely. Do not make large changes, only small one. Okay. Do not meaningfully add any placement change requests. Do not suggest any meaningful changes to placement. We want the elements all in the exact same place in the image. So only ever suggest color, style, or copy changes. Okay, that sounds good to me. All right, now we should be able to run this. This is my prompt engineering challenge more than it is anything else. So, let's gun it. Execute workflow. Going to grab the image analysis, which should be stored now. Yep. Then over here, we're going to have our change requests. Oh, you know what? We're missing something. Why are we missing something again? H Why is that? Oh, man. We are totally missing some Oh, wow. That's crazy. We were missing something this whole time. Did a reasonable job. Okay, let's now feed in the content right over here. Paste that in. Good lord. Isn't that funny? Isn't it crazy how AI can still do a pretty good job all things considered? All right. So, what is the change request that we want? Right. I'm just all over the place today. But we should be able to just like feed in a chain. Okay, this is really annoying. Should be able to feed in a change request. Then we should go change request. We should say make it orange. Add make it orange. Change text to something like upgrader systems today. Add a left click logo with a mouse pointer icon bottom right. Okay, that looks good. All right, so now we should be good. And then change request is going to be let's feed this in and say pastel hues instead. Add leftclick as the company name. flat minimalistic design. Let's do that. Okay. Use the provided company name, never the original company name. Okay, there we go. Okay, now that we're done with that, let us now continue on with our slow but gradual takeover of the world. Okay, so we just generated a bunch of images. The flow has generated something like 10 over the course of the last little bit, and I've just dumped all them in here. As you can see, what we're doing is we're taking the source material and then we are just like spinning it, right? And so, for instance, this is like kind of like a darker black one with higher contrast. This is like a much lower contrast one, but one that might look a little flat pastel. This one is like maybe a little bit different style-wise. This is same. And we just do that over and over and over again. One thing I'm noticing here is that right now we're really just iterating off the exact same font styles and then colors are different, but like the fonts are the exact same. I think it'd be really cool if we threw in like some Sarah fonts. So uh just going to jump over here to the open image model prompt and then have it vary that. So font style serif sans serif display font size font leading font tracking font color saturation brightness background color background patterns design style and colors of assets on the page swapping one asset for another. Okay. So what I'm going to do now that I've made these changes is I'm just going to rerun. Let's execute this. And uh we're now going to generate slightly different sorts of assets. And I also changed this to call it I think test two. So now we should have test twos as you could see in here. Oh, I guess they're all tests. This is test one or test. These are test twos where I added uh a couple other things. And next we should have a bunch of other test twos. But what are you going to do? Sue me. So it's generating the image right now. That's pretty cool. Um let's go rainbow mode. That sounds fun. After it's done generating the image, it's going to convert and then upload it as per usual. Okay. So we've generated one. Why don't I take a look at what that looks like? Uh, I don't actually know which one that is. So, why don't we filter by last modified top to bottom. And then why don't we also do this as a list. That way I could see what was just generated. So, what time is it right now? 11:57. So, this one should have been one of the new ones that was just generated. And actually, you know what? Anecdotally, this seems higher quality to me because we now have some noise in the background, which is cool. You guys might not be able to tell, but we actually have like a noise background, which is sweet. This one here is fantastic because we're actually moving the assets around like we have the same actors and whatnot which are cool, but we are moving assets. I should note by the way that like you know I'm using the same actors here and like you know there are going to be some potential I don't know if you want to call them copyright claims or something using the exact same actors. So you don't have to do this. You could like generate new actors. You could make that a thing. You could generate cartoons of the actors or whatever instead. Like you could totally cover your bases here. But I'm liking this because you know this is like a true spin now. you know, instead of this source, now move the text above. And I think just because of the way that the prompt is set up and the various things that we're flipping, like look at this. This is really cool. All of the results from now on are going to be pretty ideal. Not to mention, this is just one template. So when we feed in a bunch of other templates, we should get a bunch of other really interesting images. Okay, so I guess the next question is, you know, where do we go from here? We can obviously do it. So for us now, it's just a matter of scaling this up and actually running it on like a sufficiently large data set. Build an agent with a tool called that returns 100 prompt variations. I'm going to set this as a check right now. Technically, to be clear, like I'm not using the AI agent node. What I'm doing here is I'm using the message model node. I mean, an AI agent is literally just a classifier that chooses one of three or four things. We can definitely implement that functionality later, but for now, like I built something that outputs 100 potentially different variations. We're doing like five right now, but that's okay. Uh, what we're going to do now is we process 100 images of GBT1. That's good. We've added them to the result creative Google Drive. We just need some way to notify the team via an email update. And then what I'm going to do after this, I'm actually going to add an additional step here. And it's just going to be test plus iterate until prompt is perfect or as close to perfect as possible. The reason why I'm adding this step is just because I anticipate this is going to be some touch and go. We're going to have to kind of do a few additional things in order to make this like work really well. Okay. But yeah, so the thing right now we need to do obviously is do we need to generate an email. So how do we do that? The way that loops are is you know you start with the loop, you go down the loop route over and over and over again and that done route executes last. So what do we do? we actually just do a Gmail update at the very end of this. So, I'm going to go let's just draft a message. That probably makes more sense for now rather than me actually sending stuff to my email. I'm just going to draft it. And then inside of this, I'm going to say, you know, your design creative is ready. Then over here, I'm going to say, hello. And it doesn't look like we can go multi-line here, can we? Ah, yes, we can. Hello. This is a friendly update to let you know that your recent batch of design creatives is ready. Just click the just click this link. Okay, the output from the loop over items node has this web content link and that's okay. This should be the web content link I think that we need. Oh, my bad. No, this is not the web content link. We need the folder link actually. So, do we get the folder link anywhere here? Uh, I think we got it back here. I got to rename these because these are all Google Drive right now, which is not good. So, why don't I go back and just see where did we generate the folder? Download a file. Sorry, my bad. We don't actually generate the folder using this. What we do is we just search this. So, we actually have the ID of the folder right over here. Meaning, we should be able just to create our own link to the folder. So, let me see here. If I just click share, just going to go anyone with the link can uh we should probably make this editor access. So I'm going to go copy back over here. Add this link down below. Thanks. Maybe we'll go AI designer. How about that? Cool. That looks pretty good. We will add a two email. In my case, just going to be one of these emails here. And then why don't I now just execute the step, right? I mean, we got everything. It's got the output. So, we don't actually need to Yeah, we should actually be able to pin all this, shouldn't we? Let me see if I pin this. Is that going to work? No, I'll pin all these pin these two. Now, if I click execute workflow, this should just output the Gmail. At least I think it will. Oh, yeah. There you go. Oh, however, there is a problem. These are currently outputting three drafts. The reason why is because we have three done branches. So, we just need to output one. So, what I'm going to do is I'm going to go to settings. I'll say execute once. Is it the perfect way to do things? No. But it is sufficient. And now I just go back to my email. We see a design creatives is ready. If I just open this in a new tab, and now we obviously have access to all of the designs. That's pretty cool. Awesome. Okay. So, we now have all of the moving pieces to make this system great. we have something capable of taking in an image and then describing that image, converting that image into a list of possible change requests, taking those change requests and then creating, you know, 5, 10, 15, however many we define. Right now, I'm doing five just for test purposes. We then have something that takes that image as a source, uses it as a template, makes those change requests to the template, and then basically spins things, dumps it into a Google Drive, and notifies the team. So, we have the shape of the whole flow, and the shape is fantastic, and it's great. What I want to do now is I want to make this as fantastic as humanly possible. And I I'm just going to play around with it. Do some like push and pull until I get the prompt in what I think as high quality as we could get it. So, how am I going to do this? Well, I've sort of played around with the prompt logically speaking on this batch. Oh, I guess there's one other thing. We should probably change the title. I've kind of played around with the prompt a little bit here. And, you know, it's like worked pretty well with this batch. What I want to do next is I want to play around with the prompt on different types of images so that I can I don't know maybe catch some edge cases that I'm not considering because I'm only feeding in the same image. So the first thing I'm going to do is when I generate this image and then upload it, what I'm going to do is I'm just going to change this file name to now. The reason why is I'm just going to have this be let's go to string. I think we should be able to change this to a string. This is just going to be different um for every file. Right now I don't actually know if I could make this the URL, but we'll worry about that later. Like sometimes you can't have colons or other special characters in the URL, but that's okay. It's Google Drive, so maybe it'll work. Next thing I'm going to do is I'm just going to unpin all this. Unpin all this, too. Now, what I want to do is I want to go down into the source folder. So, we'll go source folder down here. And then what I'm going to do is I want to take whatever source image I'm feeding in. And then I just want to like have a couple of other ones. So, why don't I go Facebook ad creative examples again? Why don't we scroll down here? This one looks pretty cool. I mean, like I don't like how these are so low res. If I zoom in more kind of rough, man, we don't have like really high quality source. Okay. Anyway, I'm going to make this as big as possible. Then I'm going to drag this like this. Then I'm going to go to my source folder and then drop this in. And now we should have another image that we can run tests on. Little screenshot. So we can delete this old one now cuz I just want to test this on the new one. And the very first thing I'm going to do before I do anything else, I'm just going to test this on this new flow. And we're going to see if there's like any issues with the prompt that I might not know of. Okay. And I'm just feeding in image two now. So if I go back here and then I go to the destination folder. If I filter by last modified again, top down, what do we got? So this is one of them. That's cool. Restart important conversation. A little help from a therapist. So it's got the text in perfectly, which is nice. It looks like we are modifying the sand serif and the serif. And then obviously we get different logos and stuff. The text is the exact same. And I think the reason why the text is the exact same is I just forgot to update what I'm asking for change background wise. So change pastel hues instead as left click is a company named flat minimalistic design. So how cool is that? So we didn't tell it to update the text. Didn't update the text. It only does exactly what we ask for it. So what I'm going to do now is I'm going to go back and then I'm going to ask it to make some additional changes. And then maybe you know now I want to mark it leftclick a little bit more. So I'll say pastel hughes instead. Add left click as the company name. tagline should be and then let me just see build your agency with a little help from our team. Okay. And then I'm going to say flat minimalistic design. Now we should also get that written in there as well. Okay. Now let's take a peek. Build your agency with a little help from our team. Beautiful. And then we got a really cool leftclick logo down there as well. Okay. Let's take a look at that next image. Cool. This one is real. It looks like we are messing around with the textiles as well. Wow, that's a cool logo. Holy. I should probably use that as my real logo, shouldn't I? Okay, so listen. I think that we're probably reasonable quality promptwise. There is still a few things that we need to do. Obviously, we need to like attach some sort of agent to it so we could chat. What I want to do now is I just want to work on the case where there are more than one of these. And basically what I want is, you know, instead of us just running this one at a time, I want to provide the ability to run this multiple times over multiple input images. I think it already works kind of like that to be honest. Not entirely sure. We're going to try just dumping in another screenshot and then we'll see how it goes. So what's a pretty good screenshot? We could do the video never sounded so good thing. Uh should we do that again? Time for the secret sass your AOV. Listen, Netflix for learning doesn't work. This is kind of spicy. Maybe what I do is I open this image in a new tab and let me take this one. go to my source folder. I dump this in. And then what I want to do, I just want to modify this prompt a little bit so that the change is add left click as the company name. Tagline CTA should be build your agency with a little help from our team. And then we'll go instead of pastel hues, why don't we go like um maximalist colors. And then we'll go like bright remove the minimalistic design. Now we can test a different style. Now we have two images. So I'm kind of curious what's going to happen. I'm not sure. So what I'm going to do is I'm just going to execute this. And we have two items here. So this is going to download both files. This is then going to analyze going to then generate five change requests for both files. This is taking its sweet ass time. Okay. So we've now generated two. I'm just going to look at these outputs cuz I want to make sure they're fine. Go to JSON. I can actually see both them. Both them outputed. Images of man facing a woman are cool. advertising for 360 learning. Okay, cool. We then feed this into the model. Nice. Excellent. Now, this is only generating three to be clear. Is that fine? I don't remember. Did I just mention three? I think I just asked for three. That might be why. Okay, this is generating five. So I didn't actually ask like tell it how many to do. So let's do generate five changes. Change requests. No more, no less. Should be fine. Okay, I'm now going to go to destination folder. Is this the right one? Oh no, it's not the right one. Let's um swap this one more time. This should be the right one. Get started. Okay. Some issue here with the aspect ratio of the image. I don't like this. Why do we have these issues? Was the screenshot that I took a taller one? No, this one actually looks like the screenshot I took was fine. So, we should probably find a way to fix this still. Like this output is pretty wild. Um it's a little bit zoomed in, but like if this was zoomed out, it would actually look really cool, right? So, that's fine. So, this tells me we should probably update the I don't know like the size, but didn't we feed in the size 1024 x 1024? Yeah, we did. Anyway, might just be a one-off. So, I'm going to let it generate another before I make any conclusions. Okay, this is the second one. And we no longer have that issue, which is great. Build your agency with a little help from our team with a big join now button. This looks great to me. No problems whatsoever. One thing I'll ask it is to change any assets that they're just a little closer to what the change request asks for. And I'm going to do that because like logically speaking, there may be some assets on the page, maybe like a document or a receipt or something like that for some accounting company. And what we want is we want the model to be able to change that to something that is relevant to our request. Right? If we say leftclick a automation agency does X, Y, and Z, we don't necessarily want it to be a receipt unless we're talking specifically about some receipt optimization or automation thing. So your task is generate new change suggestions that will later feed an image generation model. changes you can make font style font leading font tracking swapping one asset for another that is more related to the change request make these changes reasonable want to stick with the source pretty closely don't suggest any meaningful changes to placement the placement of assets we want all elements in the exact same place in the image so only ever suggest color style asset type font type or copy changes we'll do that I'm realizing here I didn't specify the font and maybe that's another lever we could pull Just continue going up. What we see here is we're just like iterating over and over again. Got a bunch of really cool designs here. This is like one of those geometric patterns, which I like. And now from here on out, we should be outputting the new images because we're doing three at a time. Cool. This one now looks spunky as hell. I love this. I mean, you know, like it's a little bit difficult to read contrast wise and stuff, but fonts are different. Uh, you know, messaging is different. Everything here is really cool. So, hopefully it's clear at this point that we've done most of the system and this is already capable of driving a ridiculous amount of value. What I want to do now is I just want to connect this to some sort of chat flow. And the reason why is because one, I want to be able to chat with it. And two, right now we don't really have any way to put input. I'm hard coding it. Chat is just a natural way to put input. So I think what I'm going to do is I'm going to develop an a agent and then I'm going to use this workflow as like a subworkflow in a broader agent flow. And that'll just like open the door for us to do more. Now my only potential issue with this is, you know, if I make this a sub workflow, do you guys see how long it takes to execute and to finish? It does take like a very, very long time. So that may be a problem. I'm not entirely sure if that's that big of a deal because realistically things are going to be dumped into the Google Drive folder kind of over time. So maybe because this fires off a Gmail, you could make a request and then just like keep the G Drive tab open or something if you're a design team and just see them come in cuz you know, if you think about it, you don't actually have to wait for this whole flow to finish in order for you to like start getting actual value out of it, right? This one's pretty cool as well. I like how the mouse is kind of funky. That probably makes more sense to me. But anyway, let's talk about how to actually do that. So this is a flow that is sort of on its own right now, right? We're going to connect this and make this a part of a wider flow. What I'm going to do is I'm going to call this image spinner. And then in brackets, I'll put one click to 10,00 add creatives. Then what I'm going to do is I'm going to create a new workflow over here top level. And this workflow is going to be called oneclick to 10,00 add creatives agent. Okay. Now over here, I'm going to add an agent node. That's going to have a connected chat trigger. And the system messages, you are a helpful intelligent graphic design assistant. When a user feeds in an image prompt or yeah an uh an image prompt, your task is to take that image prompt and call the image spinner tool. Do not change the provided image prompt. This is dealt with change or improve the bright image run. This is dealt with inside of the tool itself. Okay. So now that we have this, we can select a chat model. Now the specific chat model we choose isn't actually that important. Just for I don't know completeness sake, I am going to choose an okay model here. Then over here where it says sampling temperature, I'm just going to keep it at 007 or 7 rather. So now we have sort of this open a chat model. And then the memory here is just going to be simple memory. We're just going to do 10. Just going to put that over here. Let's just move this. I've uh started just moving these a little bit to the right and it's not like the prettiest but it just gets it out of the way. And the specific tool call is going to be call naden workflow tool. Okay. So how do we actually connect these two now? Well right now I have a simple trigger here. What I want to do is I want to take the simple trigger and then instead I want to do execute subworkflow. So when executed by another workflow I'm going to feed this in as the input. What you can do with this is you can actually set up an input schema. So I'm going to create one and I'm going to call it image prompt and that's going to be word for word the image prompt that I feed in. This image prompt is then going to come if you think about it be right over here. So how do I get this image prompt? Um when executed by another workflow and go dollar sign here when executed by another workflow item JSON image prompt like this. Now, I can't actually access this just yet, but I should be able to momentarily. Then, under call N9 workflow tool, what I'm going to do is I will go image spinner. And here is where I feed in the uh image prompt. What I'm going to do is I'm going to allow the model to define that automatically. Okay, I just changed my whole mind. I don't like how that looks. Then, we're going to uh feed that. Actually, let's just let it auto adjust. Nice. I like that. Let's just call this image spinner. Let's call this, I don't know, memory. Let's call this open a chat model. Cool. So now when I chat with this, we should actually trigger our flow. What I'm going to do is I'm just going to remove all of the garbage here, all of the old stuff, and I'll leave the new stuff in here. And now I just want to like chat with it and see if we could actually have it generate stuff. So hey, what's going on? Hey, I'm here to help you with any graphic design needs like generating images based on your prompts or answering questions about design. How can I assist you today? Okay, let's now do make these assets. Uh, actually, I should probably feed it in. Okay, I'm just going to go maximus bright colors. Add left click as the company name instead of maximalist. What are some other design styles? Let's try something new. And let's do art design styles. Uh, maybe not art design. Let's do like graphic design styles. There we go. Just give me like the top three. Let's go grunge. That seems pretty fun. I'm going to go grunge dark colors. Tagline CTA should be start your AI agency with a little help from our team and it should be interesting. That's what I'm going to say. I'm now going to feed this in. It's going to get the prompt and then because I've told it, hey, if somebody provides you an image prompt, feed it in exactly. It's not going to call the image spinner. Now, what is going on as we speak? Well, it is now currently running the image spinner. The fact that we have yet to get any sort of problem is a good sign because if we had some sort of problem by now, it should have popped up. The image prompts probably misformatted or something like that. So, now that that's good to go, if I go back to my Google Drive, we'll see things populate live. Okay, first image is up. Let's take a peek at this. Ooh, this is grungy, scary design. Left click, start your agency with a little help from our team. It's got that background stuff and that's got a little arrow that says get the ebook with a woman. That looks similar to the initial actress but a little bit different on her laptop. Great. Nice. All right, so we just got another gen that looks pretty badass. Okay, third grungy image. Left click. Start your age and see the little help from our team. Get started. This one's badass. Probably my favorite one so far. And we generated a couple. No problem whatsoever with this, huh? I love what they did to the logo. That's so cool. Okay, so now that that's good, we should be able to scale this up pretty easily. Why don't we do one final thing? Why don't we give the model or the agent the ability for us to tell it how many things we want to generate at a time and then we'll just expose some chat endpoint so that anybody in our team can access it. It'll be like a basic nad and chat and you know it's not perfect because we don't really get feedback from the model immediately which is unfortunate. Ideally we get some sort of feedback from the model. Uh I guess we could probably change it so that it calls it every single time it generates an image. I don't really want to do that I like the idea of making it a self-contained flow. Part of the reason why I want to do this is because I'm going to make another video soon that ties all of my systems together recently. The ad scraper and then a couple other ones. Makes it like one crazy ad parasite flow where you scrape high quality performing static image ads and then just recreate them and make them a lot better. So, I think I'm going to leave it as that for now. And then yeah, what I'm going to do is I'm just going to like have a variable that allows us to modify this section right over here. Generate five change requests, no more, no less. Right? cuz now it's you know like it just outputed it generated five it's going to do the fourth and the fifth are both going to be that chick again right cool this one's really interesting I like that but um still we should generate n change requests so what am I going to do you guys remember earlier when I magically mapped this variable down at the bottom and I said change request when executed by another workflow image prompt I'm just going to copy this over then up here this really cool thing you can do so you can actually just paste this and then instead of image prompt I'll say number of samples. Okay, so now we need to define a variable called number of samples. So I'm going to go over here go number of samples. Then if you think about it logically what I can do up at my agent level. I'm just going to stop this now cuz it's obviously working and it does what it can. It's going to go to the AI agent and go to the tool call. I'm going to refresh this now. I think refresh this is over here. Not seeing anything yet actually. reset value. Maybe I need to save this. Yeah, I think I needed to save this. Sorry. Here I'm going to refresh. There we go. I'll define this as well. And then I'll say call this tool to spin an image. Pass the image prompt through directly. If no number of samples is provided in the chat then return three as a default. There you go. Okay. And now what I'm going to do is I'm going to go back to the agent and I'll say user feeds in an image prompt. Your task is to take that image prompt is to call the image spinner tool. Do not change or improve the provided image prompt. This is dealt with inside to call itself. Also make sure to ask the user for a number of samples and pass that to image spinner as well. Okay. The user does not provide number of samples. Ask them something like how many samples per image do you want? Okay. Now let's now refresh this. Not hide but refresh this. I'll say hey what's up what can you do for me thank you AI agent will tell me we can provide graphic design tasks which is really cool image prompt multiple image samples okay so now what I'm going to do first things first I'm going to write something like bold punchy cartoon like but not anime company name workflow loop decals and icons all over the and a logo in the bottom left hand corner. Experiment with serif and sand serif styles and display styles as needed. Make colors pop. Okay, I'm now going to feed this in directly and then we'll say generate three per sample. We're now going to feed into the image spinner number of samples three and then the image prompt. Bold, punchy, cartoon-like but not anime. Okay. And now Netflix for learning doesn't work. Get the ebook. We got our little loops everywhere. And then we even have a logo of workflow loop. I'd call this like a great ad. It's fantastic. Big fan. Okay, we got a couple of more. Look at this. Restart important conversations. This is now in a fundamentally different place on the image. And the style is just crazy. Netflix for learning doesn't work. Get the ebook. This is now a cartoon style. Streaming learning. Try workflow loop. Get the ebook. Workflow loop. Right. Pretty neat. Uh I'm a pretty big fan. So I'm going to leave it there. Couple other minor changes I'm going to make to this just because I don't want this to, you know, I don't really want people just to run with this on the internet. I'm going to make it change the actor slightly so that we don't run into issues where it's the exact same actor. So, if there is an actor, I'm going to make them look different. But where I'm going to do it is actually directly in the generate image using GPT1 prompt. So, I'm going to go over here and I'll say there's an actor and no style specified. Change the way they look a little so it's not the same actor. Only do this if the desired style is not a rendered cartoon or drawing version of them. So, what I'm going to do now is we're just going to test this final and then in addition to testing this final, go back to the agent here. What I'm going to do is I'm going to expose a chat endpoint. So, I'm going to go active. Then, when chat messages received, make chat publicly accessible. Hi there, I'm a graphic design agent. You drop your desired ads to emulate in this folder. go here and I'll go share edit link. Just drop your desired ads emulate in this folder. Then after your request is sent, they'll start populating in this folder. And I will grab the copy link there. Make sure to provide an image prompt with the company name and any style replacement text and any style you want as well as the number of samples per example to generate number of designs per example to generate. Happy designing. Cool. We'll leave it at that. So, what I'm going to do now is I'm just going to click this button. Grab the URL. Open this up in a new chat. It's giving us all the information here. I wish I could make this a button. You know, I probably can with Markdown, but anyway. And now it says, "I'm a graphic design agent. Just drop your desired ads to emulate in this folder. " Wonderful. We got a source folder full of a bunch of stuff. What am I going to do? I'll go Facebook ad creative. Then I'm going to click on this. Let's get one more. Okay. I'm not really liking all these. Let's go over here and then I'll just go Facebook ads. Uh under templates. Okay. Going to get all these Facebook ads here. Happy Father's Day. This marketing thing. Well, okay. I like this one. So, maybe we'll copy this. Paste that in there. Then I will go to my source folder. That's all right. Just drag that in. Let's delete these two. Let's go over here. Creators and marketers need each other like Tonic needs Jin. That's pretty funny. Um, these are very high quality, but I'm not really seeing these selling ads very well, you know. So, I think I should go back to We're go over here to this link. This is cool. I like this. Feed this in as well. Let's feed this. Now, what else we got? I like this quite a bit. Super clean. We do this. This is sick. Okay, great. Okay, nice. So, now we have a lot. Uh, what do we do now with all the images in the source folder? Well, I'm just going to go back to the destination folder, get rid of all these, and for now, why don't I just generate two variants of each. Okay. So, I'm going to go back to my chat. What I'm going to say is I want bold, punchy, cartoon-like, not anime, company name, workflow, loop, loop decals, icons all over the place. Logo in the bottom lefthand corner sample. And then I'll say put your system on autopilot stylized text. Okay. And now we're going to enter. Let's do one per sample. And now it is gone and it has sent it over to our system. First image. Put your systems on autopilot workflow loop. So not perfect. Obviously we're a little bit cut off here, but not terrible. I should note that there are some assets in this image. We had some camera and so it did just choose to replicate the camera. Uh if I had fed a little bit more information and probably would have got it a little better. Second, five ways to power up your creative workflows of generative AI workflow loop. Put your systems on autopilot. That's pretty cool. Um, I mean really the only thing here that you may have to actually change is September 21 digital. That take you two seconds. Not a very big deal at all. Number three, grow your gym with ARX. All right. So, this is a good example of an instance where it has failed. It is not actually taking in my requests. So, we're going to have to modify some things. Creatives and marketers need each other like Jen and Donic. Put your systems on autopilot with workflow loop. Man, I really love this style maximalism we're seeing. The weekend is here. Yay, it's Friday. put your systems on autopilot workflow. Oo, I really like that logo. That's cool. It's kind of like the Airbnb one, though, isn't it? All right, never mind. So, we're not using that logo, but uh hopefully you guys see just how easy it is to iterate and spin these ads. Not a fan of the maximalism personally, so I think I'm going to go back to flat minimalistic. Um, we now have generated enough diversity across images for me to feel confident making one final change, and that is to swap out the asset. So if there is an asset in the image, let's say in this case it's a gym machine, we need to swap it out with something that is specific to what we are talking about. Okay? And you know we can't include the company name in here obviously and if there's dates or something like that there are dates we need to remove them. Okay. So now I'm just going to go back to my main prompt here. Go back to the editor. Related to the request. Cool. That seems pretty solid to me. If there are other highly specific factoids uh things like dates etc you can remove them. Okay. Things that are only present things that date the ad like actual dates remove them. Make sure the change request removes them. Cool. All right, final go. Let us now run this through this whole sample of destination images again except let's do it with our new and updated ad version. Okay, we'll go CTA text and then what I'm going to do is I'll say level up your systems company name workflow loop decals style. Oh, let's go flat, minimalistic, no outlines, colorful, but light pastel. Uh, what else should we do to like spin these? H, light pastel. Let's say automated assets, i. e. a robots and loops. Am I missing anything here? Serif and sans serif fonts minimalistic. I think I already said that. Okay. Cool. I'm pretty sure that's it. I'm just going to now run this and we're going to see how this works with our new and updated prompt. All right. How many samples perish do I want? Let's just do one. We'll just do one for now. First add level up your systems. Now we have cool little robot decals with little infinity signs by them, which is really cool. Okay, the second image is up. Nice. That's really cool. I like this. Oh, and the logo again is gorgeous. Now, we do have September 21st digital event, which is not accurate, but that's okay. Assuming that we generate, I don't know, 20 of these, and we have five. That's 100 in total. Odds are we're going to be able to find one of them that we like. Well, number three, grow your gym with workflow loop. Level up your systems. Now, we have basically gone through and adjusted the machine. In this case, looks like it misinterpreted what it is that we wanted to grow. Still did a reasonably good job aside from all of that. Okay, awesome. This one's really clean. Workflow loop. Love what they did with the logo. little background gradient. And yeah, that's it. In a nutshell, we've just built a system that end to end takes as input a bunch of source ads that you guys could literally scrape anywhere you want. No problem. Just, you know, use your little screenshot tool, dump them into this folder, and then you're good to go. And then turns them into high quality assets that are branded and then hyper specific and then unique to you. So, I showed you guys how to do it all live over the course of the last hour and a half or so. I want you guys to know that at this point anybody here could do the same. It's just a matter of getting access to the right tools and then feeding in high quality enough source data. Then obviously adjusting the prompt to your own specific use case, need or style.

Hopefully you guys appreciated that video. I had a lot of fun putting that together. If anybody has any questions, drop them down below in the comments. If you guys want to see how to make real money with this stuff, selling automations and systems just like this to other businesses, then definitely check out Maker School. It's my 0 to1 90-day accountability roadmap where I guarantee you your first paying customer or your money back. It's been followed by thousands of successful people to date and I'm active in the community literally every day answering questions, recording custom videos for people and hosting calls. Aside from that, check out below. We have the template in the description. Looking forward to seeing you guys in the next video. Like, comment, subscribe, do all that fun YouTube stuff, and I will see you all later. Bye.