How I'm Using AI to Create Meta Ads

Here’s how I’m using AI to create Meta ads.

So I have a complicated relationship with AI. On one hand, there are so many things about it that I hate.

I hate the lazy and easy-to-spot AI-generated fliers. I hate when marketers use AI to replace human engagement. I barely even look at my comments on LinkedIn now because the vast majority are from AI bots.

I hate when AI is used for deception, especially when using it to depict AI-generated humans in videos and images. I hate stumbling on a blog post that is so clearly AI-generated, from the headline through the final sentence.

But I’m not anti-AI by any stretch of the imagination. I’ve used it to do so many things I never could have dreamed of doing without it.

It helps me quickly solve technical problems related to the pixel and events. It helps me create brand new landing pages instead of starting from a template like I had done for years. It helps me create tools for my private community like the Ad Briefs Library, JonBot, and a Claude MCP.

That Claude MCP, in particular, gives my members direct access into my library while using Claude. It’s all so amazing.

But one place I’ve been especially hesitant is with ad creation. I don’t want to create slop, and while I know it’s easy, more is rarely better.

But I found myself in a strange position recently. I know that the key lever for advertising success these days is the ads themselves.

Once you simplify campaign construction, specifically for sales and leads, there’s not much you need to do in the campaign and ad set. So, as I always say, the success or failure of your campaign almost always comes down to your ads.

And your ads can always be better.

I also understand that creative diversity is important, so I need to diversify in format, visuals, and messaging. And here’s where I run into a dilemma.

Believe it or not, the most painful part of Meta ads for me is the ad creation. Specifically the images and videos.

Back in the old days when my focus was on micro targeting, I admittedly created bad ads that I knew would work for my audience. But you can’t get away with creating bad ads now, and that’s multiplied with creative diversification.

I use the multimedia experience (or new creative workflow), and you can provide up to 10 images and videos. But that’s potentially far more than 10 if you create multiple aspect ratios for the same concept.

I’m most comfortable these days creating one or two ads, not at this volume and scale. So I’ve been experimenting with how I can solve this problem with AI.

I’m a huge Claude Code user. I wouldn’t label myself a power user because I know that there are people who are way more polished with it. But every time I use it to build something, I learn something new.

So I presented Claude Code with my problem, and I also explained my own concerns about using AI to create my ads.

What I’m going to describe for you is going to be a simplification of what is actually happening. Many of the details are boring and unnecessary here, and my main goal is helping you understand what I’ve done.

The first thing I did with Claude was start a rulebook for how the ads would be planned and created. Every time I made a correction or pointed out something I liked or didn’t like, the rules were updated.

Because of this, the ads I’m creating early are bound to be the worst ads I create. They should only get better as this process is refined.

Second, I’m not an expert in design or copywriting, so I had Claude assemble a panel of experts responsible for planning and execution.

The panel includes copywriters, creative directors, designers, and an auditor. Each has a defined role and each is told to push back when judging the work.

I also want the text to actually sound like it could have come from me, and that’s important. So I have it evaluate my writing to mimic my own language, words, voice, and tone.

And I have hundreds and hundreds of pages of blog posts on my website to do that.

But I don’t stop there with the copy because I want to avoid anything sounding like AI-generated text.

I feel like a detective these days because I can immediately spot commonly used words, phrases, and sentence structure, and I document it.

So I give Claude my list of slop words and phrases to avoid. A voice editor then audits every piece of copy to confirm that it mimics my voice.

I do the same thing with the visuals. One member of the panel has a single job to catch anything that looks AI-generated. It reviews the design plan before a single image is made, and then it reviews the finished images and videos again.

If you’ve used Claude, you probably know that while it’s getting better, it’s still not great with images.

But I also have a paid ChatGPT subscription, so I have Claude send its design plans and drafts and have Codex polish or generate graphics for the ads.

And even then, the panel compares what Codex made to the brief it was given, and anything wrong goes to a list of fixes.

After two rounds of that, it stops and comes back to me. I personally go through three levels of approval of messaging, visual direction, and mockups.

When it’s done, Claude Code delivers several image and video variations for a single theme, each with aspect ratios for 1x1, 4x5, 9x16, and 16x9.

All of the creative goes into a folder, and I’m given a separate file with the five primary text, five headlines, and five description options that work with the creative.

The entire process takes a couple of hours, start to finish. It’s mostly automated, but I’m still involved, and I ultimately still build the ads in Ads Manager for now.

I should note that I also have the Meta Ads Manager MCP, so Claude is able to see what I publish, and could probably publish for me if I let it.

It also pulls the results of the previous batches before starting a new one, so the next round of ads starts from what worked.

I’ve also connected my own MCP, which gives Claude direct access to my Ad Briefs Library of recommendations on strategy and approach. All of my PHC - Elite members have access to this Claude MCP.

The ads I’ve created so far aren’t perfect, but I’m actually getting better results than I was with the ads I created myself.

I also feel like each iteration is getting progressively better as I make tweaks to the rules that Claude follows.

And that’s the most important part of all of this.

You can absolutely have AI create, run, and manage the entire Meta ads process. But without some significant input from you, the execution will be generic and based on the average of advertising advice found on the internet.

It needs your input, branding, voice, and corrections to create something that could have come from you.

So here’s the bottom of the glass.

Those are the basics of what I’m doing, but this entire process will evolve as I continue to make more ads. Every batch will learn from the last, including my feedback.

If you’re like me, I get the hesitation when it comes to using AI to create your ads. Or hesitation using AI to create anything for that matter.

But know that what it’s best at is helping iterate and create at scale. When it comes to ads, this is especially important, and doing all of that manually is otherwise a potential time suck and drain on resources.

There is middle ground, and there are ways of doing this where you retain as much of your voice and brand as possible.

It may not be you, but you’re an important part of the process and can create something you’re proud of.