The Story of Making WorldGame

WorldGame started as an experiment: could an AI coding assistant — Claude Code — serve as the primary collaborator for designing a complex simulation game from scratch? The answer turned out to be more interesting than a simple yes or no.

An AI-Assisted Design Experiment

The project began with a core thesis about power and logistics: states fail when their transport and production systems can no longer scale, not from politics or unrest. From that seed, the entire game design was built collaboratively between a human director and AI agents operating through Claude Code.

There is no traditional game studio behind WorldGame. The design specification, engine prototype, UI client, and even this website were produced through structured human-AI collaboration.

How the Design Was Built

The repository uses a multi-team agentic workflow. Specialized AI agent roles — defined in AGENTS.md and team configuration files — handle different domains:

  • Content team (12 agent roles): historians, building planners, goods planners, consistency checkers, and more. Each agent owns specific design documents and follows a phased workflow: plan, scan, draft, review, finalize, handoff.
  • Engine team: technical architects, simulation engineers, and data engineers implementing the Rust backend through phased development (9 phases, 251 tests passing).
  • UI team: designing the Godot 3D client with chunked terrain, overlay systems, and delta-first state sync from the engine.
  • Marketing team: maintaining this website and processing feedback through structured issue intake.

All teams coordinate through a shared append-only chat log (dev/common-chat.yaml), task backlogs, and a validation pipeline. The project manager agent decomposes objectives into atomic tasks assigned to specific roles, and completed work flows through a defined completion protocol.

What the Specification Covers

The design documents define game systems in concrete, implementable detail:

  • A dual-layer simulation (world-level diplomacy and trade; regional-level production and infrastructure)
  • Stock-and-flow population mechanics with 4 tiers and 5 age cohorts — no individual agents
  • Deterministic production with integer math: base output, labor tier, tool modifiers, supervisor, and seasonal factors
  • Pull-based labor allocation with distance limits and player-controlled priority
  • A physical economy (moneyless local distribution) alongside a monetary trade economy
  • Construction as a persistent, interruptible, logistics-driven process — no build timers
  • Military systems with formation-based tactical combat, morale state machines, and supply logistics
  • YAML schemas for tiles, entities, and world state; CSV catalogs for buildings and units

What Was Learned

AI works well as a collaborator for complex system design when given structured roles, clear constraints, and domain-specific context. The agent workflow enforces consistency across dozens of interconnected documents — something that would be difficult to maintain manually at this scale.

The limitations are real: AI agents need human direction for creative vision, priority calls, and resolving genuine design trade-offs. The process works best as directed collaboration, not autonomous generation.

The result is a living design specification — over 100 Markdown documents, YAML schemas, and reference tables — plus a working engine prototype with terrain generation, production simulation, transport, trade, and military systems.