POKÉLLMON is the first LLM-based agent to achieve human-level performance in a tactical combat game. It leverages three key strategies: 1) Contextual reinforcement learning, which uses text-based feedback from battles to iteratively optimize its generation strategy; 2) Knowledge-enhanced generation, which utilizes external knowledge to combat hallucinations and enable the agent to act accurately and timely; 3) Self-consistent action generation, which mitigates panicking switching behavior when facing strong opponents and avoiding combat. Battles against human players demonstrate POKÉLLMON's human-level fighting ability and strategy, achieving a 49% win rate in tournament matches and a 56% win rate in invitational matches. Furthermore, it reveals vulnerabilities in human players' exhaustion strategies and deceptive tactics.