Caches reward sameness
Prompt caching works best when stable content appears at the beginning and variable content appears later. Instructions, tool schemas, reusable examples, and stable corpus material belong in the prefix. Current user questions, fresh diffs, timestamps, and run-specific IDs belong in the tail.
Two effects, one rule
This is not only a cost trick. It also reinforces good context hygiene: stable rules stay stable, live inputs stay obviously live. The cache-friendly shape and the attention-friendly shape happen to be the same shape — that is not coincidence, it is good API design.
Structure for the next hundred turns
A cache-friendly prompt is easier to reason about, easier to diff, and easier to debug when behavior changes. Designing for the next single turn is a beginner mistake; designing for the next hundred turns is the sustained-cost win.