Descriptions:
Nate Herk walks through the mechanics of prompt caching in Claude Code, showing how a clear understanding of this one system can meaningfully extend session limits and cut token costs. Drawing on his own usage dashboard, he reports caching 91 million tokens in a single day and over 300 million in a week — with cached tokens billed at just 10% of normal input pricing, effectively making long sessions far more economical.
The video explains how Claude Code’s caching architecture operates in three layers: globally cached system instructions and tool definitions, per-project items like Claude.md files and memory, and the growing conversation layer that gets reprocessed each turn. A key practical detail is the cache TTL (time to live): Claude Code subscriptions maintain a cache for one hour of inactivity, but API calls and sub-agents default to just five minutes — a difference that can silently inflate costs during complex multi-session workflows. Herk also references a quote from Thoric at Anthropic, who noted that the team runs severity alerts when cache hit rates drop too low, underscoring how central caching is to the product’s performance model.
Three habits are offered as covering 95% of use cases: avoid letting sessions sit idle past the one-hour mark, start a fresh session when switching tasks using /compact or /clear, and use a session handoff skill to preserve context cleanly across boundaries. A free token-tracking dashboard and the session handoff skill are available through a linked community.
📺 Source: Nate Herk | AI Automation · Published May 21, 2026
🏷️ Format: Deep Dive
