How to Use AI to Produce 50 Ad Variations Before Your Designer Opens Photoshop
Insights from Will Sartorius at the Foxwell Meet-Up in Lisbon
Why It Matters (for media buyers, creative strategists, founders, and marketers using AI for efficient ad creative testing)
AI isn't here to take over creative strategy; it's enhancing it. For media buyers, creative strategists, and marketers working on ad content, this is important because real progress comes from creating reliable systems that boost quality, consistency, and speed. Will Sartorius shows how AI can help teams quickly generate, test, and refine multiple creative variations, allowing designers and marketers to concentrate on strategy instead of getting bogged down in production tasks.
Will Sartorius is the brains behind Skipper, a DTC ad intelligence platform. He’s deep in the weeds with AI creative workflows, and at the March, 2026 Foxwell Meetup, he walked members through his playbook for turning a blank brief into a finished static ad with no designer needed.
Here’s the big takeaway: most AI ads look rough, and it’s not the model’s fault. It’s because people skip the setup.
Why Your AI Ads Look Like Garbage
If you jump into an AI image model and type, 'Make me an ad for [brand],' you’ll get something back, but it’s not going to be good. What’s missing is:
A brand spec card
Copy constraints
Safe zones and formatting rules
Persona targeting
A scoring system
It can be the same AI model with the same initial brief, but you'll get a totally different output when those elements are in place.
The Three Things You Actually Need
1. A brand reference card. This is your foundation. Skip it, and nothing else matters. If you’re using Claude Code, build it there and it’ll save to your project folder. If you’re using Claude chat, download it and upload to your project knowledge base. Seriously, don’t skip this.
2. A format template (markdown file). Think of it like a recipe card for each ad type. Will’s process is to take every headline ad you can find, input them into Claude, and have it break down what they all had in common. Claude will then deliver a reusable template with copy slots, word counts, image instructions, composition rules, and safe zones. You end up with a markdown file packed with style variants (Style A, B, C, all the way to F) that Claude can pull from when it’s time to generate prompts.
You don’t even need to read the template, you just need it to exist. To make your own, screenshot 5-10 ad creatives in the style you want, tell Claude to analyze them, and ask for a markdown file. If you’re using Claude Code (Opus, max thinking mode), tell it to keep going until it hits at least 1,500 words. Let it run for 15 to 60 minutes, then save whatever it outputs.
3. Scoring agents. This is where your copy gets pressure-tested. Will runs every single piece of copy through seven different agents before anything ships. Each one checks for something different: persona fit, angle, emotion, brand voice, format match, and grammar (which, by the way, AI loves to forget). Each agent gives a score from 1 to 100, and nothing moves forward unless every score is above 90. If there’s one grammar slip, or one persona that's a bit off, it’s an instant fail.
Not all agents are created equal. Copy excellence is the big one, weighing in at about 60%, with the rest as smaller factors. The agent will keep iterating on the copy automatically, sometimes running a dozen rounds with zero human input until the brief finally passes.
The Full Workflow
Start by locking in your persona, angle, and emotion up front.
Let Claude handle the copywriting.
Run it through all seven agents and let them iterate until every score is above 90.
A human reviews and approves the copy
Have Claude generate an image prompt that matches your chosen format template style.
Drop that prompt into Flux (fal.ai) or Gemini’s image model and generate.
Quick note on tools: Will goes straight to fal.ai to tap into Imagen 3 (he calls it 'Nano Banana 2'), instead of using wrappers like Higgsfield. At about eight cents per image, you’re basically just paying the model cost.
AI tools are changing fast, but the real advantage comes from knowing how to use them strategically. Inside Foxwell Founders Membership, marketers, media buyers, and creative strategists share the workflows, tools, and testing frameworks they're using to stay ahead.
What Comes Next
Once you’ve got one solid ad, it’s time for a format explosion. Upload it to Claude along with your markdown files for testimonials, before-and-afters, UGC, offer or promo statics (whatever you’ve got). Ask Claude to spin up prompts for the same copy across every format. Run those through Flux and now you’ve got 48 variations in about an hour.
Animating statics is the next big edge: Here’s how: upload your static to Claude, ask for animation ideas, create a start frame (product out) and an end frame (product in), write a prompt for the transition, and run it through a video model. As of June 2026, Veo 3.1 is your best bet.
On cloning competitor ads: Will’s got a markdown file for that, but he’s not talking about straight-up copying the competition. The real move is to spot the problem your competitor is targeting and then recreate that angle in your own brand’s voice, using your brand reference card, your product photos, and your own parameters.
Finding the Gaps in Your Strategy
Here’s another gem from Will's presentation: he built a free tool that sits on top of the Meta Ad Library API. It breaks down ads by format, persona, angle, and emotion across industries. Connect your Meta ad account and it’ll show you how your ads stack up against what’s actually working out there.
The output is a gap analysis that says: here’s where you’re overdoing it, here’s what you’re missing, and here’s the persona, angle, or emotion combo you should test next. That feeds right back into the top of your workflow.
The logic here is that with enough brands in the mix, you get a normal distribution of ad strategies across the whole market, which is more reliable than just watching a few competitors and guessing.
Final Takeaways
The prompt itself is the least important part. What actually matters is the infrastructure – the brand card, the format templates, the scoring agents, the feedback loop. Build those once and they’ll keep paying off. Skip them and you’re just making noise, only faster.
Will Sartorius is building Skipper, a DTC ad intelligence platform. He also publishes a weekly newsletter covering AI creative workflows. This post is based on his presentation at the Foxwell Meet-Up in Lisbon.
