Claude's caching trick: first send a system prompt as a placeholder, then subsequent requests are directly accelerated, almost at no cost.

AIMPACT Message, May 15 (UTC+8), practical tip to reduce API long prompt initial token generation time: pre-warming prompt cache. Send system prompts before user prompts. Claude will write it into the cache but skip generating any output. When a real user request arrives, it will directly hit the pre-warmed cache. (Source: AiHot)