Reduce prompt size without losing intent to cut costs and latency, especially for long contexts, RAG payloads, and multi-turn agents.