A To-Do App Doesn't Prove Anything

Search for portfolios built with AI-enhanced UX design and you get a wall of the same three projects. A calculator. A to-do app. A weather dashboard. Clean, competent, forgettable.

They are all safe in the same way: nothing about them has to feel like anything. A to-do app saves the item or it doesn't, and that is the whole test. There's no version of it that works but feels wrong.

Game Feel Cannot Be Prompted

So I built an arcade game. You spell words mid-flight to cast spells, which means timing, responsiveness and difficulty curve all have to land or the whole thing is unpleasant to touch.

That's the whole point of picking it. You can't describe your way to good game feel. There's no prompt for "make the jump satisfying". You build it, play it, notice it's slightly wrong, and change one number. Then again. Then again.

Spelling a word mid-flight to cast a spell. If this felt wrong everyone would know instantly, which is exactly why I picked it. Lion Wars in its own isolated component, so a refactor in one mode can't quietly break the other two.

Most of the Work Was Deleting

The model was genuinely fast at producing mechanics. It would happily give me another mode, another power-up, another system. Each one worked in isolation.

Almost all of them made the game worse, because a game is a small number of things that combine well, and a list of features is the opposite of that. Scope discipline was most of the design work, and it's the part that no amount of generation speed helps with.

Pick the Project That Can Fail

If you want to know whether someone can ship with AI, look at whether they picked something that could have come out bad. A to-do app can't really fail. A game can fail while every individual piece works.

That gap, between working and being good, is where the design job now lives, and it's exactly what a safe what a portfolio proves now project is built to avoid.

What shipped, and what I cut to get there, is in the Fire Lion case study.