PRODUCTION SCENARIO
A claims triage bot sends an 8,000-token policy manual as static system content, followed by one final block holding a timestamp and the claimant's message. The team put cache_control on that final block and sees zero cache reads across 5,000 requests a day.
Where should the cache breakpoint go?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: On the last block of the static manual, so every request shares an identical prefix up to the breakpoint
Claude's prompt caching writes exactly one cache entry at your breakpoint, a hash of the prefix ending at that block, and on later requests it looks back for prior writes rather than for stable content. A breakpoint on a block whose timestamp and message change every request never matches, so the docs say to put it at the end of the static prefix, not on the varying block; a longer TTL keeps alive an entry whose hash still never matches.