I Built an AI That Tests My Facebook Ad Creatives 24/7

Nicolas Korvessis built an automated system that generates Facebook ad variations, tests them in isolated ad sets, and scales winners every 24-48 hours using Claude and the Meta Creative Platform. The workflow pulls performance data directly into AI, which then creates new creative ideas and launches fresh tests without manual work.
Chapters
Here's what ad testing used to look like for me. I would sit down on a Monday, try to come up with 10 to 15 different ad variations, hooks, visuals, different copy angles, and then I would try to track everything in a Google Sheet kind of like this. I'd launch everything into the Ads Manager, let them run for 3 days, and then maybe a few of them would come back with enough data where I could make a right decision or scale, and the rest were wasted budgets, not the right angles, And I was spending 6 or more hours every week just on creative testing, not even on strategy or scaling. And the worst part is I was still guessing which angles to test. And now my creative testing process is fully handled with AI with a little bit of input from me, and it takes me less than an hour to do every single week across a lot of different offers. There's organized naming, clear winners, and here is exactly how the full system breaks down. And I'm gonna show you how to build the whole thing out in this video. So the system overall has 4 parts, and this is the system map. First, AI generates the creative variations like hooks, angles, and visuals, and it generates these off of a few different ways. First, just off of me feeding it winners and me giving it new ideas based off of winning hooks or winning ads. Second, I actually have my AI hooked up via MCP to things like Foreplay and other different ad spy tools that can actually pull things from competitors. Next piece in the system is a very specific campaign structure that isolates what's actually working, and I'll break that down as well. AI then analyzes the results every 24 to 48 hours. Winners get scaled, losers get replaced, and new ideas get generated from those tests. So let me walk you through each one, and I'll show you the real ad account that actually runs these and the results from them as well. Okay, part 1 is the AI creative generation. In the feedback loop. First, a note on the setup. I'm using VS Code and I've got Claude Code linked up via the extension. I'd highly recommend actually using it this way, especially if you have multiple offers, maybe multiple clients, or even multiple products under the same business. And the reason why is VS Code is great because it looks through all the previous references of stuff that you've created. So you can see here on the left-hand side over here, I've got all these different folders for all these different offers that I work on. So at any different time, if I'm working on an offer for myself or a client or for a different product or offer, it has so much information about everything here within VS Code that it can reference it. So when I say something like, hey, here's a winning ad that we had this past week for blank client, let's move forward with making some new variations, it can pull and reference all the details from that client instantaneously. So I don't have to kind of refeed it knowledge every single time and try to catch it up to speed, which is super, super useful. So make sure you use VS Code. If you want a full setup on that, I've got some of that stuff inside of our school group, and you can check that out as the first link in the description. But let me show you kind of how to make everything and the prompt to go along with it. Okay, so first is the prompt structure itself. The prompt essentially is first off framing everything as, hey, you're my Meta Ads creative strategist. I'm going to give you a winning ad or winning ad hook. And I need you to make a batch of variations using angle, hook, and style matrix. Then we'll basically just give it the winning ad, which I'll do in a moment. I'll explain what made it work, generate 12 variations, styles, rules, and then it'll give me for each variation the hook, why it works, and then the best format for each one. So what I'm going to do is I'll just go ahead and fill this in, and then I'll show you what that looks like. Now Before I do fill this in, the other way that you could feed it winning ads besides just taking screenshots and uploading it in is actually by using Foreplay, which is an AdSpy tool where you can actually link everything up via MCP into Claude Code and VS Code. And you could go and start searching for ads that you think are working or you think that are kind of getting attention. And you could do it by brands as well. So say you wanted to find ads from, Um, let's go brands here. Say like Dr. Squatch, for example. Well, you could feed all the ads from Dr. Squatch and you could really filter it here based on ads that are still running, based on the runtime being over 30 days. Usually it's a good indicator for an ad. You can see some of these 56 days, 56 days. These would all be really great examples to feed. 63 days for a static image ad. To everything. So you could hook up the Foreplay MCP to Claude Code and do it that way, or you could just do it how I'm going to do it, which I'm just going to feed it some winning examples. So let me go ahead and plug those in. Okay, so here are some of the angles that it generated for me. I could go ahead and pick a few that are my favorite. And then what we can do from here is we're going to actually send it another prompt that will actually go ahead and generate some outputs of real creatives that we can actually use. Now, honestly, the easiest way to do this is if you get a Higgsfield subscription. And if you just link up Higgsfield via MCP again to Claude Code and VS Code, that's going to be the best way to actually do this because you can use Higgsfield to make images really with a ton of different image models like Nano Banana, GPT, SeedDream, Higgsfield, pretty much everything that you would need to. Even Groq. There's a lot of stuff in here. And as long as you link the MCP up and have a good hook and a good kind of foundation and a baseline of the creative and then a good prompt, which I'm gonna show you that now and give that to you in the description. It will kind of make that happen. So same thing here. And let me go ahead and tweak this prompt a little bit. I have named this skill in my setup the image engine skill. So I'm just gonna tell it choose the best winning hooks from above. And then let me just give it some brand context. Okay, so I've gone ahead and fired off this prompt. It's the exact one that's below. The only thing I changed is I told it specifically to use the Higgsfield MCP to make these ads. And then the second thing that I did is I told it the client and the offer that I want ads to make it for. And because I'm using VS Code and it has all that material from previous conversations I've had for that client, it's just gonna go and reference everything it needs to for the client's niche, ideal target avatar, desired outcome, pain points, all that good stuff to really put everything together nicely. So we'll wait for this to generate and I'll show you what some of these outputs look like. Okay, and so I've got 3 ads that it's gone ahead and generated for me. Most of these are pretty good. Obviously we could play around with the copy, the background, and different things like that, but these are pretty solid. I didn't even give it reference images, and honestly, if you give it reference images, it's gonna produce things that are a lot better. So play with that on your own. But this is the exact system. Whenever you get them generated through Higgsfield, it's gonna send them to you as links, and then you click them, it'll open in your browser. If you would rather just download them directly from Higgsfield without that whole thing, just log into Higgsfield, go to your generations, and it'll be there. Like I said, there's more that we can do with this, which I'll be covering in future videos. I just wanted this to be a basic overview for how to generate the ads. Once you generate them, a lot of people don't set up campaigns right and don't test things correctly. So I want to give a quick high-level overview on how you should be testing things inside your Ads Manager with these new ads. So first things first, I like to have everything in one asset testing campaign because it's really important to be able to really figure out what ads actually drive results. In the future, you can do things like CBO or multiple ads per ad set. But even now with Meta Andromeda, I'm finding that it's working better to just have one ad per ad set, especially when you're testing. As far as naming conventions and whatnot, it doesn't really matter too much, just whatever you find best. What I do like to do is I like to sort the tests that we're doing by whatever the start of the week is that Monday. So this is the week of the 8th, and then it's very easy for me to see, okay, what ads did we test last week? And then I can obviously go ahead and do that. From there, we'll have 3 ad sets, 1 ad per ad set. Depending on your budget, you'll set those up, but it's super important to do it this way with those creatives so you can really identify what's working quickly. So just to recap, you can feed AI different creatives, whether it's from your own ads that have worked well or from competitor ads that you've proven out to seem to be working well based on some of the metrics that we showed earlier in the video. Use those prompts that I shared to have AI generate new hooks, new ideas, and then use those ideas, feed them into Higgsfield or some other image generation tool to generate those ads and then plug them in and test them so you can actually see the results and then just rinse and repeat. Now you can automate this a lot further if you hook up your MCP to the Ads Manager. You could have Claude be looking for those very specific campaigns. That I could feed from, find the data, and then actually execute new creative tests weekly for you on autopilot. Really cool stuff, which we'll continue to cover in future videos. But this should give you a really good starting point of being able to make tons of ad creatives, as many as you want. We just did 3 today, but you could do the exact same prompt and get 30 images. Obviously your usage is going to go up, which could get more expensive over time, but we've done that and it's totally up to you how many or how little you make. I hope you enjoyed this video. Subscribe for the next one where we'll break everything down further. And we even show you how to make video ads from scratch, which is really, really great using all AI. And if you want deeper dives, if you want more support and more access, check out the school group. It's the best school group online as far as paid ads marketing and direct response advertising. So check that out and I will see you in the next video.










