Teams often focus on infrastructure for LLM cost optimization, overlooking prompt caching. Prompt caching, which significantly reduces token costs, is frequently broken by dynamic content in system prompts. Dynamic elements like timestamps or user data ruin the cache, leading to full token prices. ProjectDiscovery successfully improved their cache hit rate by moving dynamic content to user messages, saving a lot of money. The core principle is to keep system prompts static for maximum caching benefit. Static elements like instructions and tool definitions should precede dynamic content in the prompts. Low cache read token rates indicate potential problems with prompt structure. Prioritizing correct prompt structure can drastically reduce costs compared to infrastructure changes. Many are missing out on significant savings because the system prompts contain dynamic data. Reviewing prompt structure and separating static and dynamic elements is crucial for cost-effective LLM usage. Savings are realized quickly, often improving the billing.