Most of the problems I see in software and design projects aren’t tool problems. They’re communication problems. Requirements driven by sales, discovery steamrolled or nonexistent, and the expectation that the tool will compensate for what people didn’t discuss. What makes a system work is the people — their ability to understand the problem together. AI doesn’t change that. It just makes the cost of ignoring it more visible.
The relatively low quality of AI-generated design almost always has the same root cause: poor understanding of the problem and the difficulty we have in describing what we need. This isn’t new. It predates AI and will continue even with AI, because it’s a human question, not a software one. I can barely describe the sandwich I want to the person at the counter — let alone a complex system.
I was among the professionals who adopted AI early, partly because I’d already been working with Machine Learning. My background in forest engineering gave me access to that language — we work with biological models, and biological models are almost always statistical. I learned linear regression applied to growth models and site indices. Then came ML courses applied to geoprocessing, tools like GDAL, and a frustrated ML project inside a startup I helped found. I’m currently finishing an MBA in data science. The point isn’t to list credentials — it’s that when generative AI arrived, I wasn’t someone discovering models for the first time. I was someone who already knew the limits.
Maybe that’s exactly why I consider myself AI-agnostic. I use it a lot, even before it became the gold rush it is today. I think it’s a useful tool, especially if you work at the intersection of many disciplines and need to communicate complex concepts quickly to diverse teams. But the limits exist and should be considered carefully.
When I created Gothicus brasiliensis, AI was essential for giving the project shape. But the project wasn’t born from AI. It was born from a repertoire I’d been building for years, without knowing exactly where it was going.
I’ve always loved fairy tales, legends and fables. Brazilian ones interest me more lately because I became an immigrant — and I started preferring to imagine a Brazil rather than live in its reality. I returned to reading those legends with a question: is there a legitimately Brazilian gothic? I remembered elements from pop culture, references scattered across my memory. This coincided with a political maturation — reading Frantz Fanon on colonial violence, Sabrina Fernandes’s course, the essay “Governing the Dead” that makes Brazil a profoundly eerie place to inhabit. Not just because of the ongoing colonial violence, but because of all the accumulated traumas — which overlap with my own.
From that convergence came first a Substack called Tropical Goth, which evolved into Gothicus brasiliensis: a manifesto, a bestiary, a map, a card deck, an oracle, a series of essays. The gothic genre in all its manifestations, filtered through the experience of being Brazilian outside Brazil.
When I sat down to build the site with Claude’s help, I could describe exactly what I wanted. Not because I’m good at prompts. Because I had years of research, readings, accumulated references, previous mistakes. AI didn’t create any of that. It helped me dust off the material and explore possibilities I would have taken months to prototype on my own. But without the repertoire, AI would have generated a generic site about Brazilian folklore — perhaps pretty, but empty. No point of view, no tension, no atmosphere that makes Gothicus what it is.
This isn’t exclusive to my project. It’s a pattern I observe daily while leading teams: those with repertoire use AI well. Those without generate noise — and pretty noise is still noise. The tool amplifies what you already know. If you know little, it amplifies little. If you know a lot but can’t articulate it, it helps you structure. But the raw material is yours.
My first design lesson, which I shared in the first post in this series: it was when I discovered WordArt, made a curved title for a school project thinking I was killing it — and the teacher docked my grade because she found my handmade posters more interesting. Repertoire is built by making mistakes. I’ve been making them since childhood. AI doesn’t replace that. It assumes you’ve already made enough mistakes to know what to ask for.
Throughout this series, I’ve tried to show this from different angles. That the AI aesthetic crisis is a historical pattern, not a novelty. That the bottleneck isn’t the tool, it’s the capacity for translation between languages. That it’s possible to use AI to reverse-engineer design, but only if you know what you’re looking at. That AI can be a bridge between devs and designers, but bridges need foundations on both sides.
It all converges on the same point: AI is a powerful tool that accesses, organises and amplifies knowledge. But it doesn’t generate repertoire. Repertoire is what remains when you remove the tool. It’s what you learned by making mistakes, reading, researching, working, living. AI can help you use that repertoire in ways that weren’t previously possible. But if the repertoire doesn’t exist, there’s nothing to access.
AI doesn’t replace repertoire. It accesses it. And the question that matters isn’t “which AI do you use?” — it’s “what do you know that’s worth accessing?”
This is the final post in a series on design, AI and data bias. Previous posts: The AI aesthetic crisis looks like the WordArt one, only worse · If you can’t explain it, AI won’t help · How I used AI to reverse-engineer my own design · AI as a bridge between devs and designers (without blowing up the process)