Prompt layout

Stable first. Variable last. Always.

1. Tools

Names, schemas, descriptions — frozen per agent version

Cache this

2. System

Persona, policy, long instructions

Cache this

3. Static context

Docs, few-shots, retrieved corpus that is reused

Cache this

4. History

Prior turns, appended only — never rewritten or reordered

Cache this

5. New user turn

The only tokens that should change on this request

Compute this

On Claude, place cache_control on the last block that stays identical — usually the end of history, never a timestamp or the incoming question.