Introduction

Sam Brovda works in sales, not software engineering. Still, between mid-June and mid-July, he pushed 172 commits to a side project called statback. Most of those commits were co-authored by an AI coding agent, specifically Claude Opus 4.8. The result was a daily sports trivia game built in roughly one month. This is not a tutorial on learning to code. It is a candid look at what happens when a non-technical founder ships fast with AI agents, and what breaks when real users show up.

What Happened

Statback was a daily sports trivia game modeled after Wordle and Immaculate Grid. Players received a 3x3 grid of redacted athlete cards and had to identify the player using stat lines. Everyone saw the same board each day. There were three attempts per card, optional hints that deducted points, and no timer. The game launched with baseball and quickly added NBA and NFL. Players collected cards in a digital binder, complete with rookie cards, foil parallels, and limited print runs.

The technical stack included Next.js, Supabase, Vercel, PostHog, Resend, and web push notifications. Features ranged from magic-link and Google sign-in to monthly leaderboards, email reminders, an admin panel, and a referral system.

The commit history tells the story. Day one alone included the game engine, a vintage clubhouse UI, MLB Stats API integration, and multiple sports. But speed came with cleanup. Over the next few weeks, Brovda and his AI agent rewrote hints repeatedly, adjusted difficulty five times, fixed a critical database wipe, resolved phone overheating from card animations, and iterated on sharing mechanics five different times.

Why This Matters

AI coding agents can dramatically accelerate development, but they do not replace product judgment. Brovda's experience shows that shipping fast with an agent creates real operational risks: hints that accidentally reveal answers, database architectures that collapse under load, and frontend animations that drain mobile batteries. For non-engineers and solo founders, the lesson is that launching is only the beginning. Observability, data safety, and ongoing maintenance are still human responsibilities.

Key Takeaways

  • AI writes content quickly but does not check it against what is already on screen. Hints had to be rebuilt multiple times, and one live hint even included a raw JSON artifact.
  • Instrumentation should ship before tuning. Analytics revealed that difficulty changes pushed hint usage from four per board to twelve and cut perfect scores nearly in half.
  • A failed database read treated as an empty store nearly wiped the entire database. The response included write guards, hourly snapshots, and a plan to move from a single JSON document to real tables.
  • Visual polish has performance costs. A looping CSS animation for foil cards caused phones to overheat.
  • Sharing behavior evolves. The team tested image links, card links, plain text, versioned URLs to bypass preview caching, and eventually referral codes.
  • Maintenance fatigue is real. Brovda shut statback down because every fix opened another issue, but he treated it as practice for a consumer mobile app he is now building.

Conclusion

Statback is no longer running, but the experiment was not a failure. It proved that a sales professional with no formal engineering background could ship a multi-sport trivia game in about a month using AI coding agents. It also exposed exactly where agent-generated code needs human oversight. Brovda is now applying those lessons to a new consumer mobile app. For anyone considering AI-assisted development, the takeaway is clear: move fast, but watch what the agent builds, measure what users actually do, and never assume the first version is safe to leave running.