What $800 in AI Credits Taught Me About Building

In one month I burned through $400 of Lovable.ai credits, $400 of Kilo tokens, and an incredible amount of patience.

All in the name of building a real, production-grade AI-coded app.

The First Attempt

I started in vibe coding mode with Lovable—fluid, flexible, fast. Then I shifted into AI engineering mode using Kilo and local VSCode workflows. Structure, constraints, type safety, linting, semantic tokens—the whole nine yards.

At first? Incredible. We took a prototype-grade codebase and hardened it with real engineering rigor.

But then things started to break. Docs drifted. Lint rules collided. Hotfixes created new bugs. Theme system collapsed. CSS went off the rails.

I spent two days trying to rescue it. Eventually, I gave up. Nuked the local git. Kept the repo. Walked away.

The Second Attempt

Then I went back to Lovable. But this time, I came prepared.

Bootstrap prompts with principles. Set shared variables up front. Define patterns, not just features. If I pivot, I document. If there’s drift, I log it.

Now every time I finish a feature, I ask: “If I needed to recreate this exactly—what would the prompt look like?”

Lovable answers. I save it.

feature-replay-xyz.txt, ai-drift-xyz.txt, ai-coding-convention-xyz.txt

What I Learned

This is how I’m building my AI Ops stack. One painful failure at a time. And I’m not done. But this time, I’m not starting from scratch either.

The $800 was tuition. The lesson: AI velocity without documentation is just faster technical debt. Now I document as I build, log drift as it happens, and treat every feature prompt as something I might need to replay.

Still figuring out the right balance between speed and rigor. But at least now I have receipts.