AI doesn't want to replace you. Here's how it helps me.

Figma Make is Figma's AI that generates UI from a prompt. If you haven't used it yet, the pitch is simple: describe what you want, and it builds it. What it pumps out is a solid, editable starting point that you'd otherwise spend an hour coding manually.

I started by feeding it a screenshot of an outdated news page. Then I dropped in components I'd designed for other sections of my organization's website, alongside a quick brief: generate a comprehensive news redesign using these components as the design foundation.

What came back wasn't perfect, but it was 70% of the way there, and 70% of the way there in ten minutes is a different project than starting from a blank canvas.

Then I used it again. And again. Iterative!

Soon I had a slurry of wireframes and "finished" designs that I could pull inspiration from.

I didn't stop there, though. The thing with AI is that it can only look to the past for ideas. That's great for creating an established, easily recognizable design. Where the designs fall short is in the "new" category. I'm sitting here, looking at this news page and realizing, "I wouldn't use this. I'd search for this using Perplexity..."

I love a good RAG

RAG—Retrieval Augmented Generation—is what makes tools like Perplexity work. Instead of answering from memory like a well-read friend who hasn't checked the news lately, a RAG-powered tool pulls from live sources and synthesizes them into something coherent.

And sitting there looking at my Figma mockups, I realize that's what a modern news experience should feel like. Not a feed or a library, but a librarian on demand.

Figma Make had given me a beautiful, well-structured news page. Solid grid. Good hierarchy. 2019 at its peak. It looked like every other news site.

AI is a brick, and RAG, and a wrench

AI is a multi-tool. I've seen developers around me hook themselves to tokens and skip the rounds of code review, and I end up with some nightmarish solutions. The thing about a multi-tool is that it only goes so far. You wouldn't use a muti-tool to saw a branch in half (believe me, I tried)--you'd get a saw. AI served as a research starting point, a design starting point, and even a coding starting point using Claude. These informed me about set and standard ideas—and freed up my ability to think about even stronger solutions.

Don't isolate AI as just a tool, though. LLMs serve their role well as bricks and RAGs, but when you leave the mason behind, you'll end up with a pile of bricks.