The folks at Ramp Labs created a demo where Claude Code manages an amusement park - hiring mechanics, building rides, analyzing finances, and even generating CFO reports. All through the command line.
Technically, it works like this: they wrote a CLI tool called rctctl (modeled after kubectl) for OpenRCT2, and Claude receives information about the park in the form of an ASCII grid and JSON data. No graphics - just text.
What's interesting: Claude excels at data analysis, setting prices, and placing simple objects. However, spatial reasoning is still weak - there are issues with laying paths and building roller coasters.
Why RollerCoaster Tycoon? Because the game is a "Montessori set of B2B SaaS interfaces." Staff management, customer satisfaction, financial cycles - all just like in a real business.
The main takeaway: agents work best not as a replacement for human intelligence, but as "automation of diligence" - routine monitoring and optimization, while humans focus on strategy.
🧪 Demo: https://labs.ramp.com/rct
📝 Paper: https://ramplabs.substack.com/p/ai-plays-rollercoaster-tycoon
