Prompt Caching: How AI Learns "Total Recall" and Slashes LLM Inference Costs by 90%

Summary

A detailed explainer from the easy-learn-ai project on prompt caching for large language models, based on Anthropic's best practices for Claude Code. By marking a breakpoint in a request, cached prefix computations are reused across turns, cutting input token costs by up to 90% (cached tokens cost 10%, with a 1.25x one-time write fee) and lowering time-to-first-token. The article covers why Anthropic monitors cache hit rate as infrastructure-level metric, the optimal prompt ordering (static system instructions and tool definitions first, growing chat messages last), three pitfalls that break the cache chain, and advanced techniques: Plan Mode via message injection instead of tool changes, lazy loading with lightweight tool placeholders, cache-safe forking for conversation compression, and sub-agents with isolated contexts. It frames prompt caching as constraint-driven architecture built around prefix matching.