My seven-year-old came to me with a very specific idea. A game where you design your own dragon, then navigate it through a maze, while answering multiplication questions to earn time. She knew exactly what she wanted. She wanted to choose the colour, the size, the eye shape, and whether the dragon had wings. She wanted the multiplication questions to be hard enough to be interesting but easy enough to beat the timer. She wanted to be able to play it on the iPad, and she wanted to beat her sister’s score.
None of this was on me to design. It was already designed. My job was to sit next to her and type.
You can play the result at /dragon-maze or at timtrailor-hash.github.io/dragon-maze. It took roughly sixty conversational turns with a coding assistant to build. Most of those turns were her, not me. The credit screen says “Designed and built by Rose Trailor using Claude Code” because that is literally what happened.
How the sessions went
I would sit with her at the laptop. She would say “I want the dragons to have different kinds of eyes.” I would type that into Claude. Claude would propose a way to do it. I would read Claude’s proposal back to her in plain English (“it says it will draw the eyes using curves, so we can change the shape for each kind”). She would say yes or ask for something different. The code would change. We would refresh the page and she would look at the dragon.
This loop, repeated sixty-ish times, is the whole story. The language model handled the translation between what she wanted and the code that made it happen. I handled the translation between what the model said and what a seven-year-old would understand. She handled the design.
Some of her choices were not obvious ones. She wanted the timer to count upwards, not downwards, because it felt more rewarding (“you are gaining time, not losing it”). She wanted the wrong-answer penalty to be zero, not negative, because she said penalising wrong answers made people stop trying. She wanted three answer choices per question, not four, because four felt like too many. She wanted the multiplication difficulty to be harder for seven and eight times tables because those are the ones she is learning, and easier for two, five, and ten because those are boring. The final game is tuned exactly the way she asked for.
The one thing we spent disproportionate time on was drawing the dragons themselves. She was specific about the look. The procedural drawing code (which draws each dragon mathematically rather than from a pre-drawn picture, so that every customisation combination produces something unique) ended up being hundreds of lines of Bezier curves. I reran that part of the code many more times than was strictly reasonable, because the dragons mattered to her.
What the game can do
Customise a dragon (eight colours, six eye types, three sizes, wings on or off). Every customisation produces a visually distinct dragon, drawn on the fly.
Play through ten increasingly large procedurally generated mazes. Each maze is generated fresh; no two runs are the same.
Answer multiplication questions to earn time. The questions are drawn from configurable tables; the default is skewed toward the tables she is currently learning at school.
Compare scores against other players through a shared leaderboard. It runs across devices; her cousins can play on a different iPad in a different house and the times show up side by side. One entry per player per level, keyed on the best time.
Play on a phone. The controls switch to a touch D-pad with swipe support, the layout reflows for a tall screen, the text sizes scale, and the keyboard for entering a player name works properly on iOS.
Why I am writing about it
Partly because it is a lovely thing and she is proud of it. That is the simplest reason.
But also because I think it illustrates something about the current state of coding assistants that is easy to miss when the conversation is about enterprise workflows. A seven-year-old designed and shipped a game, in a format her cousins could play, in a handful of afternoons. The bottleneck was not the code, because the model wrote the code. The bottleneck was not the design, because she knew what she wanted. The bottleneck was the translation layer between them, and the translation layer was me, sitting in the middle, reading the model’s suggestions aloud and typing what she said.
That is not a future vision. That is a thing that has already happened in my house.
The game has now been played enough times by enough friends and cousins that I cannot seriously claim the leaderboard is fair. My seven-year-old has practised more than any other player and her times on the early levels are embarrassing to compete with. If you want a shot at the top of the board, the later levels are where the field is most competitive.
Credit goes to Rose. Code is at github.com/timtrailor-hash/dragon-maze. If you are a parent with a kid who has a clear idea and you have not tried this, please do. The ratio of “how clearly they can describe what they want” to “how much of it becomes real” has never been higher.