- We need to review prompts to include some of our latest learned. - Prompt should also be better structured to optimise prefix caching (see [here](https://bentoml.com/llm/inference-optimization/prefix-caching#how-to-structure-prompts-for-maximum-cache-hits))