Why language models suddenly got good at poker without any training

Poker has long required either massive computational solvers or years of training. PokerSkill bridges the gap by feeding LLMs a structured library of expert-designed poker skills, letting a context engine filter recommendations to sensible moves. GPT-4.5 and Claude 4.6/4.7 now play within striking distance of state-of-the-art bots like Slumbot—without any game-specific training or solver queries. The insight: neither raw rules nor raw language models work alone, but together they crack a complex imperfect-information game.