Raise the hit rate

Eight habits that make hits the default.

01

Freeze the left edge

No timestamps, request IDs, localized dates, or A/B copy in the system prompt. Put those in the last user turn if you need them.

02

Stable serialization

Canonical JSON key order for tools. Some languages shuffle maps and change the prefix without changing meaning.

03

Append, never rewrite

History only grows at the end. Editing turn 2 of a 20-turn chat invalidates every token after it.

04

Clear the minimum

If you sit just under 1,024 tokens, expand static instructions. Silent skip is the usual failure mode, not an error.

05

Stay inside the TTL

A hit refreshes the clock. Quiet gaps kill the entry. Use 1-hour Claude TTL, or pre-warm before users arrive.

06

Mark the last stable block

Claude writes only at breakpoints. Automatic mode on a changing last block writes forever and never reads.

07

Warm, then fan out

Wait until the first response starts before parallel twins. A write is not visible to siblings that launched with it.

08

Pin routing if you can

Use prompt_cache_key / user when offered so related traffic lands on the replica still holding that KV prefix.