I wonder if this would be good for vibe coding / natural language for enemy AI. IE, place an enemy down and tell it: “every 3 seconds fire an arrow at the player. If the player is within 7 tiles of you, stop firing arrows, path to the player, and attack it with a sword. When your health reaches 10% run away from the player”

This misses many details, like if health is 10% but there is nowhere to go this will produce bad result where mob is just hanging around and not fighting, or if there is no line of sight then it should not fire arrows, or if it's impossible to path it should keep firing (IF there is line of sight). In a real game when you write every condition for every possible scenario I know what I would choose between a wall of text (anything in which can be misunderstood by llm) vs clear state machine

This is the goal! We're working hard to give the AI more spatial "world" awareness with bespoke decision loops