Cache-aware prompt compression

Cache-Aware Prompt Compression argues that compressing prompts without regard for provider prefix caching can erase savings from repeated long-context workloads. Why it matters: Teams using agents, RAG, or codebase assistants often reuse large stable instruction and schema prefixes. Compression should preserve cacheable structure instead of optimizing only for fewer tokens in a single request.

Try this: For one repeated long-context task, compare three runs: raw cached prefix, query-aware compression, and cache-preserving compression, then choose by total cost and output quality.