<p>Artificial intelligence (AI) has empowered materials research, enabling rapid property prediction and inverse design. While machine learning can yield materials with desired properties, defining the recipe to create a target material remains nontrivial. Synthesis involves numerous parameters, including reactant ratios, temperatures, and reaction times. Here, we leverage large language models (LLMs) to explore how variations in synthesis recipes affect material properties. Using reinforcement learning (RL), we efficiently navigate the recipe space to achieve targeted properties. This iterative loop between RL agent and LLM environment serves to emulate traditional discovery processes. As a demonstration, we apply this method to optimize cookie recipes—a familiar yet analogous problem to laboratory chemistry. Our LLM-mediated RL framework generates a recipe for cookies with a crispy exterior and soft interior, validated by experiment. This showcases the potential of AI-driven workflows to accelerate materials discovery not only with theoretical identification but also practical implementation.</p> Graphical Abstract <p>An LLM-mediated RL framework iteratively proposes and evaluates material recipes for a given target property. As proof of concept, we generate an optimized recipe for a cookie with a crispy exteror and soft interior, and validate the desired behavior experimentally</p>