Mid-Conversation System Prompts: Steering an Agent Without Breaking the Cache

Short summary

Claude API now supports mid-conversation system messages, allowing you to inject instructions after cached history without invalidating the entire prompt cache. This pattern preserves expensive cached prefixes, prevents costly reprocessing, and provides operator authority through a non-spoofable channel—safer than embedding instructions in user messages. Use it for dynamic state changes discovered mid-session that need trusted operator authority.

• •Mid-conversation system messages let you add instructions after cached history without cache invalidation
• •Preserves prompt cache integrity and prevents expensive full reprocessing on token-constrained sessions
• •Non-spoofable operator channel prevents prompt injection attacks better than user-message workarounds

Generated with AI, which can make mistakes.