Prompt optimization means removing tokens that do not change useful behavior, keeping the instructions that do, and arranging repeated prompt sections so they can be cached when the provider supports it. A shorter prompt is not automatically a better prompt.

In many applications, the user message is only a small part of the request. The larger pieces are system instructions, tool schemas, few-shot examples, retrieved context, and conversation history. Those tokens can dominate cost and latency, especially when they are repeated on every call.

Good prompt optimization is empirical. Make a smaller or more cacheable prompt, run the same evaluation set, compare cost, latency, and quality, then keep the change only if the trade-off is acceptable.