A system prompt is 4,000 tokens and contains instructions, persona definition, output format rules, and 10 few-shot examples. An architect wants to optimize for both cost and prompt effectiveness. What restructuring approach is recommended?

The architect-level approach is to first apply prompt caching (making the 4,000-token system prompt nearly free on cache hits), then systematically profile which few-shot examples are actually contributing to output quality through ablation testing. Blindly reducing examples (Option A) risks degrading the behaviors they teach without a quality baseline. Once profiling identifies low-impact examples, they can be removed with confidence. Option B (moving examples to conversation history) can work but loses the caching benefit on those examples if they must be reconstructed each session, and changes their semantic framing. Option D (splitting across calls) is architecturally incorrect — instructions must be present in each call to be effective; splitting them across calls breaks coherent instruction context.