Which prompt-construction pattern is most likely to actually hit the provider's prompt cache?

**Prompt caching** lets a provider remember the prefill work it did for the front of your prompt and reuse it on later requests. Prefill is the phase where the model reads your whole prompt and computes a key and value for every token. That work fills the KV cache, and it is the expensive part of a short-answer request. If the next request starts with the exact same text, the provider can skip recomputing those keys and values and read them from its cache instead. Because the compute was already paid for, vendors bill those cached input tokens at a steep discount, on the order of one tenth of the normal input rate in 2026. The catch is that the match must be an identical prefix starting from the very first token. Caches compare bytes from position zero forward, and the hit ends the moment the text differs. Anything you change near the top invalidates everything below it. That is why the question's correct answer puts all stable content first. The long system prompt, tool schemas, and few-shot examples form one uninterrupted block at the top, and the only thing that changes per call, the user question, sits at the bottom where it cannot break the hit above it.

**Prompt caching** stores the prefill KV state for a prompt prefix and reuses it across requests. Compute already happened, so vendors bill cached input tokens at a deep discount. In 2026 the read rate is roughly 10 percent of the normal input rate. There is a small write premium the first time a prefix is cached. You also win on latency, because skipping prefill cuts time to first token sharply on long stable prefixes. The hard constraint is prefix matching. A hit requires byte-identical content from token zero forward. The cacheable region ends at the first byte that differs. So one variable token near the top discards the cache for everything beneath it. That dictates prompt layout. Concentrate every stable element first. The system prompt, tool schemas, and few-shot examples form a single uninterrupted prefix. Per-request content goes last, after all stable text. The question's distractors fail on this rule. A leading timestamp poisons the whole prompt at position zero. Retrieved chunks in the middle cap the hit before the chunks, losing reuse on everything below. Randomizing sections destroys the prefix entirely. Only a clean stable prefix maximizes the cached region and the discount.