the prompt-structuring trick that can cut a multi-turn api bill 5-10x: put static content first, dynamic content last, so it can actually be cached
if you're building anything multi-turn and not structuring prompts for caching, your bill is probably several times higher than it needs to be. this isn't a model choice or a retrieval trick, it's purely how you order the prompt. the mechanic: put everything static (system instructions, tool definitions, few-shot examples, anything that doesn't change turn to turn) at the front