Experimenting with AI-Generated Design: Lessons from the Trenches
Over the past few months, I’ve noticed a handful of CTO friends publicly playing around with tools like Cursor to “vibe code” their way to surprisingly decent-looking web apps. It’s been fascinating to watch engineers lean into AI as a kind of collaborative creative partner.
But one thing struck me: I haven’t seen nearly as many designers experimenting in public with generative tools. So I decided to give it a go — partly out of curiosity, and partly as a fun way to promote an education project I’m running this summer.
The Project: Design Founder Summer Camp
I’m launching a 6-week course for designers interested in becoming startup founders. Over six sessions, I’ll walk folks through how to come up with investable ideas, form a founding team, raise funding, gain traction, and (hopefully) reach product/market fit.
To help promote the course, I thought I’d try using AI to generate some social media graphics. Could I prompt my way to a professional-looking banner?
Let’s find out.
Prompting the Machines
I started where many do these days: ChatGPT. I uploaded a picture of myself and gave it this simple prompt:
Create a social media promo image for “Design Founder Summer Camp” by Andy Budd. 6 weeks. Starts mid-June. $950 reduced from $1950 (show the full price slashed out). Apply at andybudd.com/designfounder. Limited places available. Use the attached image of me, cut out and with a cool, modern colour cast.
The result? Not terrible. Not great. So I asked ChatGPT to write a better prompt for itself (something it’s surprisingly good at):
Create a wide-format social media promo image for “Design Founder Summer Camp” by Andy Budd. The image should be modern, clean, and aspirational — targeted at startup founders and product designers. Use a cut-out of Andy Budd, edited to have a natural, subtle smile. Apply a contemporary colour cast… [and so on]
The visual results were passable — but my likeness kept getting weirdly distorted. Each time, ChatGPT would generate someone who looked kind of like me, but not really. No matter how I phrased it, the tool couldn’t seem to maintain the fidelity of the original image.
So I pivoted.
A New Approach: Representative Imagery
Instead of featuring myself, I decided to create a banner with representative design leaders. Here’s what I asked for:
Please create a vibrant rectangular social media graphic aimed at senior designers interested in becoming startup founders. The image should feel modern, eye-catching, and energetic.
The design brief:
Left third: Bold text layout (Design Founder Summer Camp, 6 Weeks, Starts mid-July, £750+VAT reduced from £1750, etc.)
Right two-thirds: Two cut-out images of design leaders (from an attached reference image)
Background: A vivid powder-inspired gradient of pink, orange, purple, and green — with splatter powder textures for energy.
(The actually prompt was more detailed, so this is really just to give you a flavour)
This time, things were… better. About 90% there.
But that last 10%? Something would always go wrong.
The Frustration of Near-Enough
For instance in the above image the course title is wrong. I’d ask ChatGPT to adjust the text while keeping everything else the same, but something unexpected would break. Like in this instance the pricing and URL is now wrong.
I’d fix that, but then the padding would break.
In fact padding was something that it just couldn’t get right. I’d ask for more padding. I’d ask to move the text in by 100px or 10% but nothing worked.
Instead?
- The padding would still be off.
- Or, the text would subtly change — lines dropped, £ symbols turning into $, numbers getting glitchy.
- Or the format would change (from landscape to square)
- Or, worst of all, the people in the image would inexplicably change.
Occasionally something really odd would happen. Like the background would change, the genders would change or it would suddenly turn into a cartoon for no reason.
Each small adjustment triggered a cascade of unintended changes. It became a game of prompt-whack-a-mole.
In the end, I found myself regenerating the original prompt image over and over — probably discarding 99 for every 1 that worked. The creative agility was impressive, but the lack of control made it maddening.
Some Takeaways
This wasn’t an entirely wasted experience. I got some passable banners out of it in the end, and it was fun to see what’s possible. But it wasn’t fast. And it definitely wasn’t production-ready.
If I were designing an AI-powered image generation tool, I’d consider moving away from flat raster outputs and instead generate layered files — something closer to a Figma or Photoshop file, where I can adjust individual elements (text, background, people) without breaking everything else. Or at the very least I’d use a node based UI so I could go back a few steps to tweak the inputs.
Maybe that’s where this is all heading. Or maybe the smarter move is to bring AI into existing design tools as co-pilots, rather than trying to replace them.
Either way, ChatGPT clearly isn’t ready to deliver polished visual design via written art direction alone.
On the AI Backlash
Interestingly, while I was sharing some of these experiments, a few people had a go at me for using AI to design the banners. There was a sense that as a designer, I was somehow “cheating” or devaluing the craft. That AI design was something a real designer shouldn’t be dabbling in.
I do get that.
And yet, if I’d been a CTO using AI to generate boilerplate code, nobody would have blinked. In fact, I probably would’ve been applauded for being efficient, curious, and experimental.
I found the double-standard fascinating. It seems developers are encouraged to experiment with AI. Designers? Not so much. For us, it’s still seen as a shortcut or a sell-out.
But if we don’t explore these tools, how do we shape how they evolve? How do we ensure they’re useful to us rather than just tacked onto workflows created for someone else?
For me, this was never about replacing the craft. It was about prototyping faster, exploring new visual directions, and testing the limits of what’s currently possible.
Want to see what I ended up with?
Check out the course at andybudd.com/designfounder
And if you’re a designer who’s ever thought about starting your own thing, maybe now’s your time. Spots are limited. Let’s make this summer count.
