Put Stable Instructions Before Changing Inputs
If many API requests share the same instructions, keep that stable material together. Changing text at the start of every request can prevent useful prompt-prefix reuse.
For a classification workflow, put the fixed task definition and examples before the individual ticket. Keep genuinely variable fields in the changing portion of the request. Verify the current model's caching behavior and inspect reported cached-token usage rather than assuming that similar prompts produce cache hits.
Do not move sensitive material into a shared prompt merely to improve a metric. Access boundaries and relevance still govern what belongs in a request. Compare end-to-end cost and accepted results on representative traffic; caching is an implementation detail serving that outcome, not a reason to weaken the task contract.
