We Signed Up for the World's Largest Hackathon. Here’s What Actually Happened When We Tried to Build with AI.

‍

We kept saying we wanted to explore AI-first development. We had the tools. We had the ideas. What we didn’t have was momentum.

So, we gave ourselves a deadline, with tens of thousands of people watching.

We joined the World's Largest Hackathon, organized by Bolt.new, not because we had a perfect product idea, but because we needed a reason to speed things up. And we knew we’d never get there by overplanning.

It was a nudge. A way to break inertia, stop overthinking, and just build something messy, experimental, and real. No more waiting to feel “ready.” No perfect roadmap. Just enough pressure to get moving.

And once we did? We realized that “vibe coding” with AI isn’t just about speed. It’s a completely different way of thinking, building, and learning.

Here’s why we signed up, and what it’s actually like when you’re building with AI in the wild.

‍

Why We Signed Up (Even Though We Weren’t “Ready”)

We've been thinking about adding an AI layer to our Growth Marketing Courses for a while now. We needed pressure. A public deadline. A reason to stop circling ideas and just build something.

And also? We wanted a break from treating product development like it had to be perfect. Hackathons are messy by design. That’s what makes them useful. You don't get stuck in planning. You move. You make decisions. You learn fast.

That’s especially important when you’re trying to learn new muscles, and vibe coding asks you to build a whole new set.

‍

The Hidden Difficulty of “Just Use AI”

Everyone loves to say: “Use AI, it’ll make development faster.” What they don’t say is: the tools are still raw, the workflows are unfamiliar, and it takes a lot of trial-and-error to even figure out how to start.

Here’s what we’re actually running into:

• Prompting isn’t obvious. Asking GPT-4 to “generate onboarding copy” sounds easy. But the real work is learning how to frame context, give examples, set constraints, and refine outputs. One word change can break the whole thing.
• Supabase is powerful, but unforgiving. Want to store custom business objects, manage sessions, and index embeddings? You need to understand how Supabase handles auth, row-level security, and schema management. There’s no handholding.
• Vectors and embeddings are abstract, until they aren’t. Everyone throws around “semantic search” like it’s plug-and-play. But semantic search doesn’t just look for matching keywords. It tries to match the meaning of a user’s query to relevant content, even if the exact words aren’t there.
  That means you’re actually building a new layer from scratch, one that combines GPT output, vector indexing (which turns content and queries into math), user-specific filters, and custom ranking logic to decide what’s most relevant. There’s no template for that.
• Chaining tools = building glue. Bolt.new, GPT-4, Supabase, and Zapier don’t “just work” together. You have to build the connective tissue, managing state across requests, handling fallback logic, and deciding what lives where.

Every single one of these layers is a potential trap. There's a lot of trial and error. When you’re trying to prototype fast, even a minor misstep can spiral into a two-hour rabbit hole.

‍

Why This Still Matters

All that said, we’re not discouraged. The friction is the point.

This hackathon gave us a low-risk environment to struggle in "public." It made the stakes low and the motivation high. And it reminded us that building with AI isn’t about automation, it’s about collaboration.

Vibe coding forces you to let go of the fantasy of clean roadmaps. You have to stay in learning mode. You have to work with the tools as they are, not how you wish they were.

And that’s the skill worth building.

‍

What We’re Actually Practicing

This hackathon is helping us practice things that don’t show up on the roadmap but matter way more in the long term:

• Prompt writing as a development skill
• Rapid iteration without overcommitment
• Debugging vague outputs and weird edge cases
• Designing logic that’s more probabilistic than deterministic

If you’re leading a team or building product in 2025, these skills are the foundation for working in modern stacks. Not just because they’re trendy, but because they’re effective, once you know how to wield them.

‍

What’s Next

We’re still mid-way through the hackathon. We don’t have all the answers yet, but we’re in motion.

And honestly, that’s the win.

This isn't about launching a polished product. It's about putting ourselves in the arena, learning out loud, and developing a shared intuition around how AI fits into real workflows, not just the hype cycle.

‍

Follow our journey through the World's Largest Hackathon. We’ll keep sharing the technical wins, the weird dead ends, and the things nobody tells you about building with AI. If you’ve been waiting for a perfect moment to try this stuff? This is your sign to start messy.

‍