Prompt caching reduces latency and cost for LLM applications by avoiding reprocessing of repeated content. Key techniques include structuring prompts with stable prefixes first, separating stable from dynamic components, normalizing text before caching, using content hashes for application-level caches, and caching RAG context with proper permission-scoped keys. A comparison of OpenAI, Anthropic, and Google Gemini caching implementations covers control models, write costs, read discounts (~90% off), cache lifetimes, and hit guarantees. The guide also covers caching full model responses for deterministic tasks, cache invalidation triggers, common mistakes (like accidentally changing prefixes or using weak keys on private context), and a practical implementation checklist.
