Patrick Hulin wanted to know if an LLM could reverse engineer his favorite childhood game. So he pointed Claude Code at SimTower, a 1994 elevator simulation from Maxis, and let it work. Result: towers.world, a live collaborative multiplayer clone that replicates the original's behavior. Two phases, one failure, one success. Claude fixed bugs autonomously for up to eight hours straight.
Hulin built a framework called reaper to let coding agents analyze binaries through Ghidra. Static analysis flopped. SimTower was too complex. Each room's sims have 10 to 15 possible states with complex transition functions, and the game caches everything in packed bitfields. Claude kept making premature conclusions and inventing bizarre terms like "lower-atrium band" for code it didn't grasp. "I'd say they approached a functioning simulation," Hulin wrote. "But the simulation never got to a point that was playable."
Static analysis was a dead end.
Breakthrough came when Hulin switched tactics. Dynamic analysis. He had Claude build a Unicorn emulator with mocks for all 195 Windows 3.1 API functions SimTower calls. That took about 30 minutes and was 99% correct. From there, Hulin traced state snapshots from the emulated binary and fed divergences back to Claude. Longest unattended run: eight hours. Claude fixed five parity bugs on its own.
Clean-room design protects against copyright claims, but LLMs complicate that defense. AI's black-box nature makes it hard to prove an independent audit trail. Hulin's approach, which watches external behavior rather than decompiling code, reduces some risk. But autonomous code generation means limited human oversight at the translation layer. towers.world is live anyway, letting players cooperatively build towers in a faithful recreation of a game that's otherwise stuck on obsolete hardware.