backend dev here, first game. I can code, but game design, balance and "feel" are not my thing. so for Mitts I set up a small team of claude code subagents (designer, engineer, art director, artist, playtester, marketer) and made myself the producer. the interesting part turned out not to be the code, it was the AI doing design work.

the game: you're a raccoon raiding trash cans at night. grab loot, dodge flashlights and dogs, bank it at the storm drain before dawn. the fuller the sack, the bigger the bonus, so it's always "just one more bin". the games is free, no ads, no mandatory signup, it is available for desktop and phone:

https://playmitts.com/?r=aigamedev

https://reddit.com/link/1xhxk6b/video/yhuk0xo5c9yh1/player

how the AI does the design side:

‑ the designer agent owns a design doc with every rule and tuning number. any gameplay change has to update it in the same commit, so the agents (who start each task with zero memory) always read the current rules.

‑ the game rules live in one deterministic js file that also runs in node. so the designer doesn't guess at balance: bots with different play styles (careful 2-items-a-trip, greedy full-sack, etc) play thousands of nights in seconds and it reads the clear rates.

‑ a playtester agent can also play the real game in a browser, move by move, and read the screen as data. useful for bugs, too expensive in tokens to do all the time.

‑ after launch a tiny anonymous counter (no cookies) tells us where real players clear, fail or quit.

where it got interesting, real players vs the sim:

‑ launch day: nobody cleared night 4, 0 out of 6. the designer ran the sim and found the quota jumped too fast for careful players (only about half of the "2 items a trip" bots cleared it). it capped the early quotas, the sim went to \~77%, it shipped the next day. since then 4 out of 8 real tries cleared it. better, still the hardest night, and i'm ok with that.

‑ someone left the tutorial right after reaching the goal, because there was nothing to do but wait for dawn. the design fix (end the tutorial once you're home with the goal, plus a "done for tonight" button and auto-ending nights when nothing is left to grab) was live about an hour after i saw the number.

‑ almost half of desktop players left the tutorial without ever moving. on desktop the "how to move" prompt was just tiny WASD keys, easy to miss in a dark yard. now it's bigger, shows arrows too and says what to do. early numbers look better but it's only a handful of players.

what the AI was bad at:

‑ it designs for the bots. the sim says a night is fair, real players still quit, and the counter is what helps to catch this.

‑ taste. it gave me two redesigns of the raccoon's face and i rejected both. the best idea came from my family: they loved that he "gets fat" as the sack fills, so now he does.

‑ it sounds sure when it's guessing. i have to ask "is that from data or a guess?" a lot.

art, since people ask: everything in play is drawn with canvas code, no sprites. the menu portraits are made with gemini.

happy to talk about the sim setup or the agent setup. and if you play, tell me where you quit.