AI drafts the prototype. I own the flow.

Pictures were not enough

AI models got smarter. At work, a short prompt now ran until the job was done. It looked good, and it worked.

My company, OOZOU, supports AI and pays for the subscriptions. I love trying things.

Then we were asked: no static pictures. Show a real prototype, on a phone.

I taught it from zero

I was the one who started it. Then I taught my colleagues, from zero:

• 💻Claude Code in the terminal.
• 📱The installs to open the phone Simulator.
• 🔌The Figma MCP, so the AI can read the Figma file.
• 🏷️Layers and files prepared so the AI can rebuild the design. (Why names matter: note 06.)

We finished one project that way. Then the next.

AI only? It did not work

The next project started with a Claude Code demo, and the plan was AI only.

It did not work.

An app has a lot of images. AI could not make them back then.

And our prompting was weak. It was our first time working this way.

Figma first

So: design in Figma first. Connect the MCP. The AI reads Figma, then builds.

Cropping images, trying layouts, making variants: all faster in Figma than waiting for an AI preview.

Figma first, then AI builds.

Then the code got messy

Coding the prototype took months.

Sometimes I wondered: why code this much? A clickable Figma prototype might have been enough.

The code was messy, too. Designers had no programming background, so it was built page by page, not from components.

So I organised it:

• 🧩Template pages. One page, reused with many different data sets.
• 🗂️Data per page. Each page keeps its own data.
• 🐙GitHub. I taught the team that too.

I have worked this way ever since.

We hit many problems. We got through them.