Do Your AI Prototypes Always Look Like Drafts?

Last week, I asked AI to build a customer management interface, and it looked completely generic. Empty placeholders, fake data, layout like every tutorial demo—I was too embarrassed to even send it to a client.

I got stuck here too—I used to think the AI just wasn't smart enough, but switching tools didn't help. Then I realized the problem wasn't the AI; it was that my prompts were too thin.

What is the 3-Layer Feeding Method?

Ravi Mehta (former Tinder Product Lead) demonstrated a 3-layer feeding method in a livestream: the Function Layer defines what I want, the Visual Layer shows AI design sketches, and the Data Layer feeds in real data. He built a music app in Claude live—first time with only functional requirements, the output looked generic; second time adding a Figma sketch (a wireframe from design software, or just a photo), the layout improved but the data was still fake; third time, he fed in a structured list of artist names and album names (a format AI can read directly), and the resulting interface was already ready for a prototype demo.

Replicate Cost Today

Money: $0 (Claude or ChatGPT free tier is enough)

Time: 20-40 minutes for a trial run

Technical barrier: Just chatting with AI, no code needed

First step: I open my AI chat window and list the functional requirements—like "homepage needs a search bar, recommendation list, user avatar"

I also messed this up: at first, I crammed all three layers into one message, and the AI got confused. Later I found doing it layer by layer and merging them at the end works much better.

Advice by Stage

Just starting out? I found getting the Function Layer right is enough. The more specifically I write what I want, the less generic the AI output. I wouldn't worry about the Data Layer yet.

Have 1-2 clients? I started adding the Visual Layer. Just sketch something, snap a photo, and toss it to the AI; it understands the layout way better. Even a pen-and-paper sketch works.

Scaling up? We use all three layers, especially the Data Layer. I organize my existing client data into a list and feed it in, so the prototype is something we can actually show to clients or investors.

We don't all need this method. If I'm just asking AI to write an email, I wouldn't complicate it. But next time we need a product prototype or client demo, I'll try feeding it one more layer.