What an AI-Coded Sports Trivia Game Reveals About Shipping Fast
- Sam Brovda
- 521 reads
A month of building statback showed how AI coding agents speed up product work—and why developers still need to own data safety, quality and maintenance.

AI coding agents can turn a product idea into a working app quickly. They can also turn early assumptions into bugs that only appear when people use the product. A month spent building and then shutting down statback, a daily sports trivia game, offers a practical view of both sides: the agent accelerated implementation, while product judgment and maintenance stayed with its human builder.
Sam Brovda, who comes from sales rather than engineering, made 172 commits between June 18 and July 16. Claude Opus 4.8 was listed as co-author on 155 of them. The pace produced a game spanning three sports, user accounts, leaderboards and collectible cards. It also produced repeated rounds of fixes for clues, difficulty, storage, performance and sharing.
A daily game with a collecting layer
Each day, players received the same 3-by-3 grid of nine baseball, basketball or football player cards. Names were hidden, along with birthplace and initially team; the visible stat lines served as clues. Players had three guesses per card, with optional hints that reduced their score. The format drew on Wordle and Immaculate Grid, but added a binder that collected each correctly identified player.
That binder gave the game a longer-term loop. Cards could be rookie editions or foil parallels, with serial numbers and limited print runs. A Gold parallel, for example, was limited to 10 copies per daily drop. The app ran on Next.js, Supabase and Vercel, with PostHog for analytics, Resend for email and web push for notifications.
The feature list grew rapidly. On its first day, commits covered the core game, a vintage clubhouse interface, a binder animation, MLB data, a monthly leaderboard and account sign-in. NBA and NFL support, holographic finishes and real jersey numbers followed in the same burst. That pace made a broad prototype possible; it also meant that some product details were still being worked out after launch.
Clues needed more than fluent text
Hints were one of the most persistent quality problems. Early clues sometimes repeated information already visible on a card or included unsupported accolades. Brovda first replaced stat-based hints with trivia generated by Claude, covering 888 of the 930 players in the pool. One generated line even shipped with a stray JSON fragment at its beginning.
The next attempt grounded clues in Wikipedia and Wikidata, producing verified hints for 914 players. Yet grounding facts in reference material did not solve the product problem. By July 16, team names appeared on every card, while scouting-report hints named the player’s team in 60% of entries. The facts could be accurate and still give away information the interface had already supplied.
That distinction matters for any AI-generated feature. A model can produce plausible text at scale, but it won't reliably check whether that text fits the current screen, game state or user task. Someone has to validate the clue in context, including after the interface changes.
Difficulty changes need stable boards and measurement
The game’s difficulty mix changed at least five times. Brovda moved through several distributions of easy, medium and hard cards, experimented with a guaranteed “white whale,” and adjusted the lineup after observing player behavior. One recalibration changed boards that players had already started. He reverted that change, then added logic to pin a daily lineup the first time it was served, preserving both live and archived boards.
Analytics helped expose when a later adjustment overshot. On July 3, PostHog showed the average number of hints used per board rising from about four to about 12, while perfect 9-for-9 results fell from roughly 41% to 20%. Brovda responded by shifting back to a mix weighted toward modern players.
The sequence points to a useful order of work: instrument the experience before tuning it. Without usage data, a change in difficulty can feel like a reasonable adjustment while making the game frustrating in practice. Stable daily boards matter too; otherwise, players may compare results from different versions of the same puzzle.
The data shortcut became a safety and speed problem
A June 24 commit addressed a serious storage failure. The app treated a failed database read as an empty store, then allowed fallback logic to write empty defaults over existing data. Within 40 minutes, follow-up commits added a safeguard against writes that cut core data in half and hourly snapshots stored in a 48-slot ring. Another commit outlined a move away from one large JSON document toward structured tables.
The same storage design also hurt performance. Two days later, a commit tied slow gameplay to a roughly 1 MB JSON document where every write competed for a single database row. The architecture had helped get the prototype running, but it concentrated both risk and contention. For teams using coding agents, the key question isn't only whether a feature works; it's what happens when reads fail, writes overlap or the dataset grows.
Polish and sharing had real device and platform costs
A rare-card foil effect caused phones to overheat. Its animation loop changed a CSS variable in a way that triggered constant full-page repaints. The visual treatment looked appealing, but its implementation kept the device working continuously. Card flips also ran backward when opened, and the first fix didn't fully resolve the behavior.
Sharing took several iterations. The first options included image previews and individual card links; the design then shifted to readable text with a link after recipients preferred text over an image. Link previews on X and iMessage remained cached for days, so the app added a version number to share URLs to refresh them. Later, referral codes were added, offering both the existing player and a new friend a “Scout” foil card when the friend joined.
Fast creation still leaves the hard calls to people
Within about a month, statback had three sports, sign-in, leaderboards, collecting, email reminders, an admin panel and referrals. The speed was real, but so was the cleanup. Brovda says he shut the project down because maintaining it became a headache for a side project: each repair seemed to uncover another issue.
His experience suggests that AI agents can compress implementation time without removing the work of watching a product in use. The human still has to decide whether clues are useful, whether difficulty feels fair, whether data survives failure and whether visual effects behave on phones. Statback became a practice run for Brovda’s next effort, a consumer mobile app. The durable lesson is to pair fast builds with early analytics, clear data safeguards and time reserved for maintenance.