I build governed AI decision systems for regulated utilities, and I research world models and agentic AI.

Prompt caching reuses work the model already did on a repeated prefix. The match is exact, token by token, from the very first token. I have seen teams start the system prompt with the current time. It felt harmless and it kept answers aware of the date. It also changed the prefix on every single request. So every call missed the cache and paid full price. A request ID, a user name or a shuffled tool list does the same. Even reordered keys in tool schemas can break the match. The fix is layout, not tuning. Put stable content first: instructions, tool definitions, reference documents. Put anything that changes last, near the user turn. Then measure the cache hit rate per route, not the bill alone. Track time to first token beside it, because a miss shows up there first. Treat a falling hit rate as a regression and alert on it. The trade-off is small. The model still sees the date, just later in the prompt.