In this tutorial, Nicolas Korvessis demonstrates a production pipeline that reduced the time to generate over 100 Meta ad creatives for ten clients from 20+ hours to under two hours. He walks through a four‑step workflow built around Claude (via Claude Code in VS Code), custom skill files, Nano Banana for asset rendering, and optional post‑processing with Pletor. The process begins by feeding client context (brand guides, hook banks, ad structure guides) into Claude, which first generates 20 ad angles (hooks, bodies, CTAs). These angles are paired with visual direction via skill files that map copy themes to image concepts. Nano Banana then renders static and background assets from the detailed prompts. For further refinement, Pletor can ingest the assets and angles, run them through additional models (e.g., Gemini 3.1), and produce multiple visual variations. The system is orchestrated through VS Code and a Slack bot that triggers the pipeline and stores results in Google Drive. Korvessis shares the custom skill files, explains hook bank creation, visual pairing rules, and demonstrates the end‑to‑end output, emphasizing scalability, volume‑driven testing, and the business impact of cutting creative production time while maintaining high ad performance. He concludes with a community invitation for further resources.
Chapters
Last week, my marketing company shipped over 100 Meta ad creatives across 10 different clients. We spent under 2 hours actually making them, and a year ago, this would've taken us 20 hours or more. I'm gonna show you the full stack— Claude, Nano Banana, and a couple of custom skills I built in VS Code— so you'll have everything you need to run it yourself by the end of this video. Let's get right into it. Meta ads today need volume to have success, and 20 creatives or more every single week is the new floor. It's no longer the ceiling. We spend over 6 figures every single month on Meta ads for our clients. And this is the system that we actually use to generate winning creatives, not a theoretical one. Most AI creative tutorials or even different AI creative platforms or tools are really just basic. This is a super realistic workflow that can be used at scale, which is why it'll scale past 10 creatives in one flow. So just to break down exactly what this looks like. What we'll do is we have this 4-step process. First, we feed client context into the model. We do this inside of VS Code with Claude with some custom skills that we've made. You can check those out by checking the link in the description, and I'll break those down for you here in a moment. From there, it actually generates the 20 angles first, not the creatives. We get the hook, the body, and the call to action. We separate it in this way with a few different styles as well. And I'll show you what that looks like. From there, we have Nano Banana render the static and background assets. And from there, if everything is good to go, we're good. But if we want to take it a step further, we'll use another app called Pletor, which actually takes those assets and takes these angles, takes them a step further and can even generate them with other AI models. Because then you end up having something like this, which is really cool, where we'll have new inputs here that we generated with Claude in VS Code. I have a new prompt that Gemini 3.1 will run through, and then from there I can run it through as many different AI models as I want with a few different variations per image. So this one workflow here generated me 4, 5, 6, 7, 8, 9-ish, 10, 12, 13 creatives, give or take. And that's not even counting the ones that I did originally. Inside of Nano Banana. So I wanna show you the system from start to finish. Let's get right into it. Okay, so I have everything set up inside of VS Code using Claude Code, and what this allows for is some really good file structure, especially if you have multiple offers or multiple clients that you're running this for. What I'm gonna do is I'm just gonna spin off a prompt down here, asking it to make me some new ad creative briefs. And then while it does that, let me show you what's actually inside these skill files. So these skills are pretty basic, but they're detailed. I'll basically have a few different options here. I have the primary skill here, which is this image prompt engine and this pipeline. I make it very clear that, hey, you're not actually making the image, you're making the ad concept with visual direction for future AI to expand more so downstream. It can pull different ad copy briefs from different clients. It can check clients' brand guides if it has any. It can read something called a hook bank, which I think is a super powerful asset every single company should have. If you don't have like a Google Sheet or a Google Doc with a ton of hooks, just a ton of super simple one-liners as the callout or as just the hook, you're missing out because you can be testing those hooks, iterating on them, and then actually tracking what's working. Then we have this ad structure guide that I put together where basically I told it, hey, for a winning image ad, they usually consist of a callout, a hook or the meat, and then some kind of CTA. So follow this structure every single time. I've got different hook formulas that I put in here. So like how to achieve outcome without this hated thing. If you are this avatar doing this behavior, read this. Do you want this result in this timeframe? So on and so forth. And it actually gave it these 4 different hook sources based off of how it can actually do the copy, right? So pain points, dream outcomes, Unique mechanisms and points, your own proofs. And then if the hook bank does have winners, riff off of those angles. So it's all about just building these skill files in a way where if you reverse engineer how you make winning ads or what the winning ads that you have, how they're structured, you can make a super easy skill file like this. Like I said, you can also click the link below and you'll get access to these, but I think it's good to make something that's based off of your winners. The last 2 skill files that I'll make is the actual angle to visual pairing. So I basically gave it instructions here that says, hey, once we have the ad angle and all that kind of stuff, I want you to put it together in like a good-looking visual approach so it actually looks okay. So if there's a pain, put before state imagery. If there's an outcome, show confidence, space control, achievement, clean tones, etc., right? If there's comparison, split screen, side by side, all that kind of stuff. I mean, I'm just describing based off the angle, what should the visual kind of look like based off of that. Now what's really cool as well is I actually hooked all this up to work inside of Slack, which is really cool. So what I actually ended up doing is I made a bot in Slack where I'll put the client name, I'll tell it what I'm looking for, and then it'll make all these images based off of that same engine. So I technically don't even have to leave Slack. It's not showing the images here, but I can actually open up a Google Drive where I have everything housed, and it'll actually store creatives for my clients or for my offers, and it'll drop them right in here. And some of these crush it and do very, very well. This is a dating offer that we're running, and it's doing great. So that's essentially that high-level workflow inside of VS Code. Now let me show you what it looks like once we get that end result and then how to use it. Okay, so I gave some details of an offer, and now my VS Code bot is basically giving me some prompts here. So what I'm gonna do is I'm just gonna pull a few of these. I'll copy this one. And now we have a few options. We can just go directly into Nano Banana and we can actually have it generate these images cuz you can see like a really strong prompt here that it has. We'll send that through, it'll do its thing. The other option is we can actually jump into Pletor, which just kind of maps this out visually. And then we can actually map some of this out here. You could essentially do the same thing just directly in nanobanana or even in VS Code and kind of build it out and string it together with nanobanana. But Pletor is nice because you can visually see it. Then I've got a separate prompt in here that basically tells it how to handle the inputs that I'm giving it. And what's really cool is you could also give visual references too, which is nice if you're trying to test a variation of a winner or something like that. So let's see, this one's thinking here. Cool. This one's not bad. Super easy creative, looks pretty clean. And I'm sure this will work well for this dating offer that I'm talking about. And you could essentially run this whole thing on a loop once you have more than enough creatives to get you these 20 angles. Like I said, if you're doing it in Pleitor or something similar, or even if you just stitch this up inside of Claude Code through Slack, You could generate a ton of angles like this Slack workflow that I showed you. Every time I prompted it and gave it all the information it needs, it would generate, I think, 10 or 5 concepts and then 8 images, right? So 4 concepts, 2 versions per concept, and it would work really, really great. We'd get all those feedbacks and all those creatives and it would crush it. And this is a really easy way to just take the idea and the concept and turn it into a strong ad creative and just test. Meta is so big on testing nowadays. You have to be pumping out volume. And even if the ad creatives are not perfect, they're going to help you guide your direction on what you should be making next to help you get the best results in your paid ads. So if you like this video, let me know down below. Let me know if there's anything else I can cover to make this better. I'm not selling a course or anything on this, but I do have a community that gives you more resources to make this easier for you to follow. So if you're interested, click the first link in the description, join us in there, start making more money from your paid ads and getting more success all while working less. I'll see you in the next one.
