I asked Claude to help me design a logo, and it was a disaster

I have been paying for the $20 Claude Pro plan for a couple of months now, and I continue to do so. When it comes to vibe-coding personal tools, it’s well worth the entry fee. I’ve used it to build my own custom screenshot-framing tool for articles; I also have a Claude Code automation that pulls in interesting links and pitches from my inbox (so I don’t have to go through it). That part just works. So I got ambitious.

We are currently in the process of setting up my wife’s dental practice, and the branding and logo are a big part of it. We have a designer on board, but she was unavailable for a week or two, a crucial time when we had to figure out the outdoor signage situation. Here in India, that means dealing with board designers, technicians, and printers, all with their own unique requirements and limitations.

So I thought, why don’t I use Claude to make this process simple? Why don’t I just use Claude to make the signage? I already had the wordmark ready and drawings for the logo. I have a UX design diploma from a decade ago. As Clarkson says, how hard could it be?

Claude loves making tools

I uploaded the wordmark file in SVG format, provided a rough sketch of the tooth-based logo I wanted (incorporating a “T” motif), and included a detailed design prompt. The first attempt was way off, so I told Claude to thin out some bits and change some curves. And after 20% of my 5-hour usage limit and 10 minutes, I had something decent in my hands. Claude gave me four different ideas; we picked one and continued working on it. After an hour or so, even my wife was impressed with the results.

When I did this, Opus 5 and Fable 5 had recently come out. And I had free Fable 5 credits to use. So I spared no expense and quickly depleted all of my token reserves in the back-and-forth with the edits.

At one point, Claude itself was fed up. Both my wife and I have a keen eye for design (in hindsight, this is why I shouldn’t have taken up this project), and we kept asking Claude to change the size of certain elements of the tooth. Make the roots longer, shorten the valley between the teeth, change the crown’s curvature, and so on.

After a couple of tries, Claude just gave up and made a slider-based tool for us. Again, quite impressive at first glance. You change a slider to the distance of the root valley, and the length of the “T” motif in the logo. We used the slider to dial it in, and felt like kings. We even asked Claude to create a mockup of what the signage would look like, using our wordmark. Happy with the colors and the look, we sent it off to our designer and the signage person.

It’s all fine until you zoom in

Claude generated a very pretty-looking mockup of the signage with the colors and text we wanted. It even created the corresponding digital canvas to match the size of the signage board so there won’t be any weird stretching issues when it is printed.

The moment I sent the mockup file, the issues started cropping up. The size of the font won’t do. It needs to be bigger, thicker. And for the life of me, Claude cannot generate a thicker font that looks the same as the original one. Even after throwing tokens, the generated wordmark would lose the font’s sharp characteristics, rounding off the edges.

Claude also gave me instructions on how to deal with the signage technicians and the kinds of designs we can go for (CNC cutouts vs. 3D embossed lettering). But the technical dimensions it provided for each design were just way off. And because this is an AI, it was mighty confident that what it said was the gospel.

It only got worse when my designer opened the file. While the logo looked fine as a small icon, zoomed in, it was filled with jagged edges. And even after two days of prompting, it just couldn’t smooth out the curve at the crown of the tooth. Passable for an Instagram logo, but impossible to print out at a four-foot height. In true AI fashion, the more I prompted, the longer the context window got, the worse the results.

AI design just can’t hold up in the real world

I was just about to lose my mind at this point, serving as a middleman between a technician and an AI model. I even tried to redo the whole thing. Am I the AI’s assistant now? I thought, maybe if I ask Claude to generate the entire thing on its own, it can then easily make the changes. That was a horrible idea. The things it generated wouldn’t get a passing grade at a first-year college class.

Thankfully, my designer was back in town at the time, and we rushed to sit with her. In about an hour, she talked to the technician and understood the details that needed to be changed. She made the font thicker while still keeping the edgy look (manually adjusting the font edges as needed). With my wife’s help, she found anatomical drawings of teeth, traced them digitally, and created the smoothed-out, non-jagged logo we were finally happy with.

When it comes to AI use, discernment plays a big role. I don’t use it for any sort of writing because, to me, writing is a creative process where I want complete ownership. With this experiment, I learned that it’s the same for design.

At least right now, Claude AI isn’t ready to generate designs that can be used in the real world and actually look good (there are plenty of bad AI-generated hoardings out there, and you only need 0.5 seconds to see they were made with ChatGPT). Would I use Claude to design a landing page for the clinic’s website? Yes. I might even use it to generate a slide deck. I lost two days’ worth of productive, billable hours, and learned a valuable lesson. For anything that goes into the real world, it’s best to involve real people in the process.