How to use Claude to balance faction game AI

The developer of Caleb's Hollow used Claude Code to iterate on goal-based faction players through a headless simulator. The first large set of AI-versus-AI matches did not produce useful balance: players could remain passive, and tweaking their parameters alone did not address the underlying game problems.

The human first corrected core balance issues. Then two agents received opposite faction objectives and permission to change the player code. They repeatedly ran simulations, adjusted goal weights and introduced objectives that made the factions engage rather than merely survive.

The simulator could export the state of each turn as a GIF. That gave the developer a way to inspect a suspicious game and distinguish a real strategy from a misleading win count. The reported iteration eventually approached even wins while both factions actively contested the game.

The outcome is a more credible set of simulated players and a concrete balance-testing workflow, according to the devlog. It does not demonstrate universal balance or replace human judgment of the rules. The important distinction in the account is that the developer fixed the game before tuning agents, and checked visible match behavior alongside aggregate results. An even score alone would have concealed the earlier passive play.

Prompt
Suggested prompt (written by Agents In Use, not the author's)

Source: itch.io devlog · twistedweasel