Post from Field: Generation of Code with LLMs

Recently, I came across an insightful post by Jon Turow on AI Agent Infrastructure, which resonated with me—especially the part on code generation, an area we’ve been experimenting with at CIVVIA . I wanted to share some quick learnings and gather feedback from practitioners in the field!

I tried Create.xyz, which I first read about in Jon’s write-up. My goal was to build an app for a (fictional) Recovery Center offering various services.

• First attempt: The generated screens weren’t properly connected, and the UX didn’t allow me to fix them. I deleted the project and provided a more detailed prompt.
• Second attempt: Better results! This time, the workflow made more sense—a user could sign in, state their distress level and pain points, see recommended services (e.g., IV therapy), view a schedule, and later report results. Create.xyz emitted enough code for the frontend and a PostgreSQL database (speed of db creation is phenomenal) but navigation between screens remained an issue. I could only run them individually in ‘demo’ mode.
• Biggest annoyance: A persistent ad (or notification?) about a single sign-on feature, which links to an X (Twitter) post. Dismissing it doesn’t help—it keeps coming back!

Next, I tried Bolt.new, which Jon also suggested.

• Better UX (dark mode is a plus!)
• Generated a lot of code, and it looked solid
• But… it ran into a compile issue and never got the app to run, so I couldn’t fully evaluate its effectiveness
• Stay tuned as I will experiment more

At Civvia, we’re experimenting with AI-generated frontend code using Cursor with Claude as the LLM. The biggest advantage? It works inside an IDE. We can give it a UX (screen by screen) (Designs created in Figma), provide a set of rules for generating code, and it generates structured code exactly how we like it:

/src  
  ├── Features  
  ├── Components  
  ├── FE Code  - > Calls to backend  APIs
        

So far, progress is strong, and we’ll share more once we connect the backend with web application.

Curious—has anyone else been experimenting with AI-generated frontend development? What’s working (or not) for you?