Sitemap

Prototyping AI with AI

4 min readJun 15, 2025

It’s somewhat meta. Many of us in tech are in some form of AI game. We use and build for it at 514. I use it for daily life. More often now in Design — and with high fidelity prototyping.

Press enter or click to view image in full size

Story

Prototyping with AI can be pretty quick, frustrating and also rewarding. In a few prompts, you can go from zero to “hey, this kind of works” — pretty quickly. For internal tools, POC’s and various prototypes it’s pretty solid, especially using v0, Lovable, Gemini or Replit.

Press enter or click to view image in full size
Internal Chat I made

Vercel v0, Lovable, Replit

I’m always hunting. LLM powered AI tools like V0, Lovable, Cursor and Replit make it easy to prototype beyond Figma. They’re great for getting to something real, quickly. If you work with Vercel’s hosting service, it makes sense to give that a solid try. Figma recently released Figma Make, but I won’t be covering that here.

Vercel v0 Benefits:

  • Vercel, Shadcn and Tailwind support
  • Visual, chat and code editor
  • Cross platform support
  • Pretty good at Figma and other integrations
Press enter or click to view image in full size
Lovable

Problems, Prototypes, Proofs

Most of the time, you don’t know exactly what you need until you see it working. V0 helps you skip the overthinking stage. Want to build a bot that answers common design, engineering or product questions? V0 isn’t perfect and doesn’t respond well to everything but its good within the Vercel and Shadcn ecosystem. I decided to try and do this while killing a few birds:

1. Better understand designing for Humans, LLMs, Generative Content

2. Prototype something useful while learning (a Company Bot while I was Out of Office in this case).

3. Explore Vercel v0 as we use it for deployment vs Loveable or Replit.

Press enter or click to view image in full size
Vercel’s V0

Context Is King

The difference between a chatbot that answers and one that understands is context. With V0, you can provide docs and various knowledge to give your agents memory and awareness. Markdown (vs let’s say JSON), in particular, works great here — clean, readable, and rich enough to carry nuance. Markdown appears to work the best with prompt results. Even with provided and structured content, some models struggle to return the right information. You can also integrate with other services like Neon, Supabase and Grok. MCP’s are also great ways to inject context.

For this, it’s basically an LLM creating an LLM, sorta. Or at least the feeling of. I organized content and prompts by Company, Product, Design, and Engineering — with sub categories like ‘Assets, Flowcharts and Latest Figmas’. Coupled with freeform typing, prompt suggest and a repository of all prompts. You can also write rules to handle specific scenarios you want your LLM to follow eg ‘Referencing a certain doc or link at a specific time like GSAP or Framer Motion’.

Models matter

The smarter the model the better the result. Claude Sonnet and Gemini have been pretty snappy and more helpful than ChatGPT lately. Be sure to check versions and use Agent and or Auto modes in Cursor for additional automation.

Stay organized

On top of context, a well labeled Figma imported can help the cause. Extreme specificity in art direction and or letting AI recompose your thoughts can often help return much better results.

Press enter or click to view image in full size
Rules I made

Deploy and Fork what works

One cool benefit of building with these tools is how easily you can clone functionality if you like the baseline. Once you’ve structured your flows and context models well, forking isn’t just reuse — it’s faster prototyping. Try out features and if they work, duplicate/deploy them again and again.

Press enter or click to view image in full size

POCs vs Real Products

That said, V0 excels at prototypes, not production. Without real-time APIs or deterministic control, chatbots built this way hit limits in reliability and latency. They’re good for POCs, demos, and internal trials — but real productization will still require engineering and integration beyond the UI. Still, as a tool for shaping what’s possible, V0 is a leap forward.

Press enter or click to view image in full size

Getting Started

With free and paid plans it’s easy to try it out. Looking to build a prototype? Below are a few jumping off points. Feel free to take a look at what we’ve been up to as well.

How 514 uses AI

We build frameworks, AI tools and infrastructure to help you rapidly productize or operationalize your data. Working with Data? Try our suite of tools out below:

--

--

Alex
Alex

Written by Alex

FAANG & Startup Alum. Case Studies, Articles & Rants.