Why 99% of vibe-coded MVPs look bad — and our playbook

Everyone has access to AI agents. Everyone has access to AI design tools. Everyone can ship something. Yet 99% of MVPs look bad, feel bad, and stick to the purple gradients.

We build MVPs before the product for our signed clients. Here's the playbook we use at our studio.

1. Use GPT to engineer context, not to give answers

Take the idea and use an LLM to brainstorm deeper, add the missing context, and explain why that card should be rounded-sm and slate-50 — not some random value. Giving enough context makes or breaks your MVP.

Rome wasn't built in a day; an MVP can't be built in one prompt. I'm not a big fan of long PRDs and spec sheets, but here they work like magic. (Use any LLM you like — my personal favourite is ChatGPT.)

2. If you don't have taste, your agent won't either

Use Gemini Nano Banana, Google Stitch or ChatGPT image generation as design agents — for UI sketches, not final designs. Give them references, colour palettes you've found, and components or images you've made. I personally screenshot other websites and steal their UI. 🥰

AI-generated slop means purple gradients, rounded borders and aniyaayathuku padding — you can smell it a mile away. Use Pinterest or Dribbble for inspiration, then think about how to recreate it.

UI isn't artificial intelligence. It's human intelligence.

3. Feed Claude

Remember the word context? Claude is as good as an entry-level developer with all the right skills — and juniors only build what's needed with a senior's guidance.

Give your agent one task at a time. Ask it to do five things at once and it'll fail. Ask it one thing with screenshot references, clear explanations, supporting docs and clear rules, and it'll build exactly what you expected.

If you're pre-MVP

• Frontend — React + TypeScript + Tailwind
• Backend — Supabase or Convex
• Auth — Clerk
• Deploy — Vercel or Netlify for launch; Hetzner or Contabo for CPU-heavy loads; RunPod for GPU work

This will get you to the door — it won't make loyal customers. If you're a founder serious about building a product to solve a real problem, let's talk.