Imagine an AI that not only masters the icy strategy of curling, but can also explain its moves like a seasoned pro. That's the promise of RL-LLM-DT, a new method that combines the learning prowess of reinforcement learning with the reasoning and coding skills of large language models (LLMs). Traditionally, building AI for games like curling involved complex decision trees or reinforcement learning, each with its own drawbacks. Decision trees require extensive human tweaking, and reinforcement learning models can be unpredictable against new opponents. RL-LLM-DT offers a clever solution. It starts with a basic decision tree and then uses reinforcement learning to find its weaknesses by pitting it against a learning AI opponent. When the decision tree loses, the game data is fed to an LLM 'critic' which analyzes the failures and suggests improvements to the decision tree's strategy. Then, another LLM 'coder' translates the updated strategy into Python code for the curling AI. This loop repeats, refining the curling AI's strategy with every iteration. Researchers tested this approach by submitting their curling AI to the Jidi competition platform, where it climbed the ranks, ultimately claiming the top spot among 34 competing AIs. This success demonstrates the power of combining LLMs and reinforcement learning. The LLM's ability to generate and refine decision trees, coupled with reinforcement learning's ability to find flaws, creates a robust, adaptable AI. This method has the potential to not only revolutionize game AI, but also to improve strategic decision-making in other fields. The explainable nature of decision trees and the knowledge-infusion from LLMs open up exciting possibilities for building more transparent and effective AI systems.