I built a creative team inside ChatGPT Work that researches our competitors, writes the headlines, makes the images and launches the ads on Meta. Then, every Monday morning, it reads how those ads did and plans the next batch. That weekly review is what makes it a self-improving AI ad creative system, and it’s why each new round of ads starts from what the last one learned. Here’s how the whole thing fits together, and how to build your own.

Why an “AI ad generator” isn’t enough

Search for an AI ad generator and you’ll find plenty of tools that turn a prompt into a finished ad. That’s handy when you need one image in a hurry. The trouble starts with the second ad, because a single ad can’t tell you whether it works, so you’re back to guessing.

Paid social runs on volume. You put a lot of ads up against each other, move budget to the ones that win, cut the rest, and build the next round on what you found out. That’s a lot of making, launching and reading, week after week, and it’s exactly the work this system takes on.

The system, end to end

The system runs in four stages, which I’ll call Inputs, Create, Launch and Learn. You can see all four in the map below.

Inputs is the groundwork: Research into what already sells in your market, a brief for the campaign you’re running, and your brand guidelines. Create is where the ads get made, static images first and video later. Launch puts those ads live on Meta. And Learn reads how each ad performed and uses that to plan the next round.

Notice that the loop skips Inputs on the way round. Your research and brand guidelines still hold, so Learn hands its findings straight back to Create, in two forms: A brief for more static ads in the direction that’s working, and scripts for video ads built on the winning angles. The next batch goes live, Learn reads it, and the cycle starts again.

I run all of it inside ChatGPT Work, but the same setup works in Claude Cowork or any other agent. The one job the agent hands off is generating images and video, which goes to Higgsfield. (Higgsfield sponsored the video this article is based on. The costs I quote are what I actually paid.)

Inputs: research, brief and brand

Of the four stages, this is the one that most decides whether your ads are any good, because everything the system makes later is built from it.

Start with research into what already works on the platform you’re advertising on. If you’ve run ads before, your own results are the best place to look. If you haven’t, look at your competitors. The Meta Ad Library is made for this: It’s a free, public record of every ad running on Facebook and Instagram. ChatGPT Work can go through it for you, clicking into competitors’ ads, following each one to its landing page and building a research dashboard while you get on with something else. Mine came back with almost 70 observations about how our market sells.

Next comes the campaign brief, which is simply what you’re promoting. Ours was Black Friday for the 9x Pro membership. Pro has several parts, so I made a separate set of ads for each one. That way the results also show which part of the offer people respond to most. The first set was for the AI Operator course.

Last, point it at your brand guidelines. That’s what keeps every headline and image recognisably yours.

Create: statics first, then video

The Create stage makes two kinds of ad, static images and video, and it always makes the statics first.

The reason is cost. Statics are cheap and quick to make, so they’re the place to find out which message lands before you put money into video. Pigeon Digital makes the case well: Prove the message on statics, then take the winner to video. So Create starts with headlines, moves on to images, and leaves video until there’s data to brief it.

For headlines, ChatGPT Work writes a batch using a handful of proven copywriting frameworks, and you choose the ones worth testing. (I’ve written up the frameworks in the piece on Facebook ad copy.) For the images, it hands over to Higgsfield, which gives it a choice of image models. I used GPT Image 2.5 Sunburst. You attach reference images, describe the look you’re after, and it comes back with a set of on-brand options. The Higgsfield MCP guide walks through that step in full.

One habit will save you a lot of credits. Once you’ve picked your headlines and images, the combinations multiply quickly. My picks came to 12 headlines, eight images and two background colours, each in two aspect ratios, which is hundreds of ads. So I have it make a single variation first, and only when that looks right does it make the rest.

Launch: live on Meta

With a batch of statics ready, the Launch stage puts them live. Meta Ads isn’t one of ChatGPT Work’s built-in plugins, so you add it as a custom server using Meta’s own Meta Ads AI connectors. Once it’s connected, ChatGPT Work builds the campaign, the ad set and the ads to Meta’s best practices, at whatever budget you give it. I started with six ads at 50 euros a day. The Meta Ads MCP guide has the full setup.

Learn: the weekly loop that improves the ads

Learn is the stage that turns a faster way of making ads into a system that gets better on its own.

Once your ads have been live long enough to gather some data, two things happen. Meta’s algorithm starts moving spend to the ads that perform, so the winners pick up budget without you touching anything. Meanwhile, ChatGPT Work pulls the ad-level results through the Meta Ads connector, sends them to BigQuery for the heavier analysis, and works out which angles, headlines and images are driving results. It then briefs the next round in two ways: More static ads in the direction that’s working, and video scripts built on the winning angles.

You don’t have to do any of this by hand. I set it up once as a scheduled task that runs every Monday at 9am. It pulls the active campaigns, analyses the ad-level data, checks our ad repository for the next thing to test, and posts the next six ads to our team’s Slack channel along with what it found. How often you run yours depends on how much you’re spending: Daily if you’re getting plenty of impressions, every couple of weeks if you’re just starting out.

A word of caution, because “self-improving” gets thrown around a lot. The loop is real and it runs by itself. In the video, though, the Learn stage worked from an earlier campaign’s data, because the Black Friday ads had only just gone live. What the system reliably does is make and test the next round. Whether that lifts your results, and by how much, is what the loop is there to find out.

Video from the winners

Video is the second thing Create makes, and Learn writes its brief. By the time you get here, the statics have shown you which messages and angles work, so you can put the expensive format behind ideas that have already proved themselves.

The same setup handles it. You record a rough clip, or have ChatGPT Work write you a script, and it sends the footage to Higgsfield, which cleans it up, restyles it and finishes it with a video model such as Seedance 2.5. The full before-and-after, from a phone recording I made on a sick day to an ad ready for Meta, is in the piece on AI video ads. The habit from the statics carries over too: Render a few seconds on the model you’re considering before you commit to the whole video, because video is where the money goes.

What you need to build it

The tool stack to run this ad system is lean. It consists of just three tools, at least two of which you’re probably already using regularly:

• ChatGPT Work, the agent that runs everything: The research, the headlines, the campaign build and the weekly analysis. The same system works in Claude Cowork if that’s where your team lives.

• Higgsfield, connected through its MCP, for the image and video generation, with a choice of models behind it. It runs on credits, and its pricing starts at around $15 a month.

• The Meta Ads connector, for launching campaigns and pulling back the performance data that Learn works from.

Most of the cost is in generation, and it stays modest until you get to video. Static images come out at around 9 cents each, while a finished video ad runs to tens of dollars. That gap is the whole reason the system tests on statics and saves video for last.

Where to start

Build it one stage at a time. Begin with Inputs: Point ChatGPT Work at the Meta Ad Library, let it research your market, and get one batch of static ads live. Add the weekly Learn loop once there’s data worth reading, and bring in video when the statics have shown you what works.

Everything I used to build this, the skills, the prompts and the markdown files for every stage, is free in the 9x community, where our operators share the workflows they actually run. Grab it, point it at one real campaign and let the first loop run.

Common questions

What is an AI ad creative system? It’s a loop. An AI ad creative system uses an AI agent to research what works, make a batch of ads, launch them, read the results and build the next round on what it learned, so the ads get better each cycle. An AI ad generator, by comparison, makes a single ad and stops.

Is this different from an AI ad generator? Yes. An AI ad generator turns a prompt into one finished ad. An AI ad creative system wraps that generation in a process: Research, brief, create, launch, learn, repeat. The generator becomes one step inside the system.

Do the ads really improve on their own? The loop runs on its own: It reads the performance data, picks the winners and briefs the next round on a schedule. What it can’t promise is a specific lift, because that depends on your offer, your audience and your creative. Think of it as a machine that reliably makes and tests the next round, and let the results tell you what’s working.

What tools do I need to build one? Three: An AI agent like ChatGPT Work or Claude Cowork to run it, Higgsfield for image and video generation, and the Meta Ads connector to launch campaigns and pull back the data. Everything else is the prompts and skills that tie them together.

Where should a beginner start? With the research and a single batch of static ads. Get one campaign live, add the weekly analysis once there’s data to read, and only move to video once the statics have shown you which messages work.