I’m building a build-planner for an unreleased game — in public, mistakes included.
It started as a side wiki for theorycrafters. 190+ commits and three months later it has accounts, saved builds and daily traffic — and I’m still shipping on it every day. Claude wrote confident, well-formatted strategy commentary next to the game’s real data. It was also wrong. I caught it before it shipped and pulled every AI-authored opinion, keeping only the transcribed truth. A security gate was quietly failing open instead of closed — fixed. Google flagged parts of the site as low-value, so I gated the thin pages instead of writing more of them. And when the highest-traffic pages started buckling, I traded raw speed for a cache that actually knows what’s safe to reuse.
The actual edge was never the AI. I curated real, specific in-game data that most people can’t cleanly get their hands on, then used AI to process and structure it faster than I ever could by hand. The AI didn’t invent the advantage — it scaled one I already had.
It’s one of more than 70 things I’ve shipped or prototyped this way — a live chat-driven mini-game for Twitch with its own backend, a tool that assigns hundreds of students to classes in seconds, a local-business site that’s generating real customers right now, a bilingual app waitlist that converted from day one. Different problems, same core method, different tool for each job.
