How I actually use an AI assistant to build a small browser game

I've been writing software for about twenty-five years, and over the last year I've shifted a lot of my hobby work onto AI coding assistants. Not because they're better than me — they're not, at anything I actually care about — but because they let me get a playable prototype in front of one of my kids in an evening instead of over a weekend. That speed is the thing. The faster the loop, the more games we ship, the more they learn.

I want to walk through one specific example so this post is useful instead of vague. Last month one of my kids said, "can you make a game where you see a circle and then you have to draw it from memory?" That became Perfect Circle. Here's what actually happened.

The initial prompt

This is what I typed into the assistant, almost verbatim. I keep these in a notebook because I learn more from re-reading my prompts than from re-reading the code:

"Make a new game where the goal is to draw a perfect circle. The player sees a glowing outline for a second or two, then it disappears and they have ten seconds to recreate it from memory using mouse or touch. When they finish, overlay the real circle on top of theirs in a different colour so they can see the difference. Score them on how close every point on their stroke was to the real circle. Five tiers — F up to S. Be objective; actually measure the deviation, don't fudge it."

What the assistant got right

The first draft worked end-to-end. It rendered the guide circle with a fade-in, it captured the stroke as an array of (x, y) points, it scored by computing the distance from each point to the nearest point on the true circle, and it printed an S-through-F grade. Maybe sixty percent of the gameplay logic was usable as-is. That's the magic — getting from nothing to something playable in one shot.

What I had to fix by hand

Almost everything that made the game feel good. The first draft sampled the user's stroke at every mouse event, which on a high-refresh-rate monitor meant the scoring was dominated by points so close together they barely moved. I had to resample the stroke at fixed intervals. The radius detection used the bounding box of the player's stroke, which meant if you drew a small loop and then a stray line away, the algorithm thought your circle was huge. I switched to the median distance from the centroid. The grading curve was also way too generous — every first-timer was getting an A. I had to look at twenty or thirty real attempts (mine and the kids') and recalibrate.

None of those fixes were hard. Each was five or ten lines. But they're the kind of judgement calls the assistant can't make for you, because it doesn't know what "feels right" means for a game. It's never played anything.

The lesson, if there is one

Treat the AI like a fast junior who can type. It gets you to a draft that would otherwise take hours. Then you play the game thirty times in a row, write down what's wrong with how it feels, and you fix those things by hand. The fixing-by-hand part is where my kids actually learn what "good code" looks like, because they can see the before and after.

I'm not interested in the "AI will replace programmers" debate. The way I use it is the way a carpenter uses a power sander. I still have to know what the surface should feel like.

If you want to see how this particular one turned out, play Perfect Circle here. Beating S-tier is harder than it looks.

— Chris