01 / DECIDE
Architecture decision
Use it for repeated large context where cache behavior can reduce latency and usage.
Keep durable instructions and reusable context stable while placing volatile user data later in the request. Map input, output, state, tools, and side effects as one observable path, then name the owner of design stable prefixes for prompt caching.
02 / OPERATE
Three-step field procedure
- 01
Lock the contract
Write down the caller, data classification, success condition, timeout, cancellation, and ownership. Use it for repeated large context where cache behavior can reduce latency and usage.
- 02
Run one bounded path
Build one end-to-end path with correlation, typed state, and reversible failure handling. Keep durable instructions and reusable context stable while placing volatile user data later in the request.
- 03
Verify the signal
Turn acceptance into a repeatable fixture, contract test, or browser test. Telemetry compares cached tokens, latency, cost, and task quality without logging private content.
03 / BOUND
Operational boundary
Do not duplicate sensitive content merely to improve cache hits; retention and authorization remain primary.
04 / PROVE
Acceptance evidence
Telemetry compares cached tokens, latency, cost, and task quality without logging private content.
SOURCE / HTTP
Reproducible source probe
curl -fsSI 'https://docs.x.ai/developers/advanced-api-usage/prompt-caching' | sed -n '1,5p'