Fire Lion

Every "I built this with AI" portfolio piece is a calculator or a to-do app. Those can't fail in a way you'd notice, which is exactly why people pick them.

A game can. If a jump feels wrong, everyone knows immediately and nobody needs the mechanics explained to them. So I built one. Game feel is the thing a prompt can't hand you.

The rule I set for myself was that the core loop had to be playable before a single mechanic went on top of it. One tap, one lion, nothing else. If that wasn't fun, no amount of spells or modes was going to rescue it.

The spelling came after, and only because it survived that test. You spell a word mid-flight and it casts. It works because tapping already felt good, not because it's a clever idea.

Fire Lion escalates and gets tense. Lion Wars is slower and strategic. Cub Mode is the low-stakes one you go to when you've been losing.

Each one lives in its own isolated component, with its own state and its own audio. That's a defensive decision, not an architectural preference. I was building at speed with AI, often late, and a refactor that reaches across a shared file breaks things you won't notice until someone else finds them. Isolation costs duplication and buys me the ability to change one mode without holding the other two in my head.

AI scaffolded the Supabase schemas, the particle systems, and refactors across thirty-odd files at a time. That's real work and it did it faster than I would have.

Then there's gravity, tap impulse, and the difficulty curve. I tuned every one of those by hand, on a phone, over hundreds of runs, because there's no way to ask a model whether a jump feels right. It'll answer, and the answer means nothing.

Producing features got cheap. Deciding which ones survive didn't.

Daily missions. Streaks. A Wordle-style daily challenge. A social proof counter. A forge upgrade screen. Mod gating. All of it built, all of it shipped, all of it removed.

None of those were bad ideas on paper, and that's the trap. They were retention mechanics bolted onto a game that hadn't earned retention yet, and each one put another screen between opening the thing and playing it.

The game got better every time I cut one, and it stayed true long enough that I stopped treating deletion as cleanup.

The palette is the only part of this game a token file could hold. Colour carries the state because there is no time to read anything. The gravity and the tap feel were hand-tuned on a phone.

It's live at firelion.me and it's finished enough to hand to someone without explaining anything first.

AI can build a game in about a week. That is the least interesting thing here. Somebody still has to hold the line on what stays in, and no model was going to do that for me.