Every time a new model comes out, I start something new, or drag an old project forward, to see what's changed. These are the projects where Claude did most of the typing, in the order the models arrived.
Commits from this stretch just say "Claude".
When Claude Code first came out, I used it to let my kids make a bunch of games, mostly in pygame. This arcade memorializes them.
Over a dozen small games my kids made by talking to Claude: artillery, Flappy Bird for two, snake, a platformer, monster slicing, pirates, racing, Spacewar, Orcs Must Die and line infantry, plus a guest game, Quarterback Simulator, from the Bullock Bros. Each one is ported to the browser and has its own arcade cabinet.
I wanted to explore two things with AI: imperfect information, and hierarchical communication between AIs. The hex grid never really worked, with all sorts of weird bugs, and we had so much trouble getting the engine itself going that we barely got to the generals. It might be fun to hand this back to a future model and see what it can do.
A WWII-style hex battle simulation meant for LLM generals: a major general and five commanders per side, each seeing only part of the field and passing orders down the chain. A Python engine writes battle logs and a Phaser app replays them.
The first time model names show up in my commits.
My son wanted a giant game, and a giant game needs a lot of NPCs. I wondered whether they could run themselves, like the LLM town simulations I'd seen, with the settlers talking to each other. It started as an experiment in how far Claude could get. When Jev came out in September 2026, this was the obvious place to try it. I never tuned it enough to work well, but Jev was fast, smart and cheap enough to run NPCs in real time, and that was pretty cool.
Six LLM settlers and one wolf on a procedurally generated island. Each settler has a daily planner, a second planner that runs all day and handles speech, and Jev for moment-to-moment activity.
RL Fighter went from first commit to self-play training in one weekend.
I worked as a data scientist from 2018 until ChatGPT came out, and reinforcement learning was something I never got to play with as much as I wanted. This was just to scratch that itch. Someday it'll probably end up in the arcade.
A space battle that runs live in your browser, looping through team deathmatches, carrier battles and convoy escorts. One neural network, trained by PPO self-play, flies every fighter by thruster and trigger, while scripts handle the carriers and freighters.
Big enough to stand up a whole platform: Vercel, Cloud Run, Supabase, CI.
Civil War General, itself built on Panzer General, was one of my favorite games as a kid. I wanted to see whether Claude could do more than write a game like that: could it also make the art, in a disciplined way? It started with Opus 4.8 and PixelLab, an image generator that specializes in pixel art. Then Fable 5 came out, and this is the project where things Opus couldn't work out on its own, Fable started working out. It built a pipeline that paints terrain tiles to match the concept art, checks them automatically and has me pick the winners. It's definitely unfinished, and I don't know if I'll go back to it.
A turn-based hex wargame where both sides plan their orders, then everything executes at once. Painted terrain tiles, a computer opponent and a replay of every turn.
SpaceSim and Bob's World picked back up; 18BBY and the droid voices started.
SpaceSim started years before AI: I'm interested in microeconomics and emergent behavior, and wanted to simulate a whole economy with as few shortcuts as possible. Space makes it easy to see. Planets are near-perfect markets, and ships hauling goods between them are the friction. I wrote it by hand, first in Python, which was too slow, then in Kotlin, where I trained a reinforcement-learning trader to act in its own self-interest. When Claude Code arrived it was the natural place to try it, and I've come back with every major model since. Fable was the first time an AI made my jaw drop on something this complex: it found design decisions I'd made long ago, explained clearly why they were causing problems I'd been stuck on for years, got my buy-in, and fixed them.
An interplanetary economy: about a hundred individuals per planet producing and consuming, a market maker, and ships trying to turn a profit along the star lanes. It's a place to explore arbitrage, rational action, and where money comes from. Runs locally only.
The curse of 18xx games is that there are so many versions, and playing one makes you want to make one. After a recent game I thought a Star Wars version would be cool, with hyperspace lanes instead of railroads, so we vibed it up quickly. Claude built the game and then playtested it with bots, though I'm not yet sure how well that worked. It's playable, but certainly not balanced, and probably not fun yet. I'd like to go back to it.
An 18xx-style game of hyperspace commerce, set 18 years before the Battle of Yavin. Corporations chart a shared hyperspace network and earn revenue by pairing the planets they can reach. Hotseat play in the browser.
I'm excited about the Microduck, a tiny two-legged robot from Pollen Robotics, and I'm expecting one on my doorstep in December or January. Customizing it is kind of the point, so I've started exploring how. This is the first experiment: a voice for it.
Procedural space-opera droid speech. Describe the intent and emotion, or type a line and a small LLM picks them, and get chirps, whistles and trills with no samples. The same request always makes the same sound, with timing a robot can move to.
439 commits in two weeks.
Opus 5.5 was doing so well that I started thinking about games I'd written off as too ambitious to vibe-code. One I've always wanted puts you on the throne, making one decision at a time and watching it ripple through the world. Reigns does that beautifully, but under its storytelling the simulation is just four sliders. Crew keeps Reigns' two-choice cards and puts a much deeper simulation underneath. Opus wrote the code, and a very small LLM narrates, making up the story as it goes. It doesn't work very well yet, but it's interesting to watch sometimes.
You're the captain of a small freighter. The crew run themselves: systems break and get fixed, cargo moves between six ports, and off watch the crew play cards and hold grudges. People bring you decisions, and you answer each with one of two choices.
During the relay we could see our place in our own group, but not where we sat in the whole pack. I wanted to see that bell curve, then watch it move through the race, and it turned out to be easy to put a map under it. I had the idea just before my leg, and it was done before our last runner crossed the line, about an hour and a half later.
Every runner in the 2026 Akron Marathon, half marathon and team relay, moving around the course on their chip split times. Scrub the race clock, watch the field spread out, and find a bib.