In this work, we aim to further make progress on learning policies with LLMs. The new bottleneck is not the learning approach itself anymore but a representational bottleneck. Therefore, we want to increase the language expressiveness towards numeric planning and other more expressive planning formalisms. Moreover, we also want to tackle to what extend we can learn reusable control knowledge that transfers to different tasks, similar as functions in standard programming. This requires language extensions in the form of contracts, which place additional path constraints on the execution of such control knowledge.