Your application needs to maintain a consistent persona for Claude across all user sessions. The persona definition is 3,000 tokens and never changes. The system also includes dynamic user context injected per-request. What API design minimizes token costs while maintaining persona consistency?

The prompt caching approach is correct because prompt caching is the correct API mechanism for static, reused content like a persona definition. By placing the persona in the system prompt and marking it with a cache control breakpoint, the first call writes the cache, billed at the base input rate, and all subsequent calls read from cache at approximately 10% of the input token cost. For high-volume deployments, this provides about a 90% cost reduction on the persona prefix. The dynamic user context is injected after the cached prefix and is billed normally. The user-turn message approach is wrong because user-turn messages are not automatically cached by default — prompt caching requires explicit cache control breakpoint configuration, and system prompts are the appropriate location for persona definitions, not user turns. The fine-tuning approach is wrong because Anthropic does not offer a self-service fine-tuning API for Claude models that would allow baking persona definitions into model weights; this feature does not exist as described. The abbreviation and shorthand compression approach is wrong because abbreviations and shorthand in a persona definition would likely change Claude's behavior in subtle ways, such as tone, vocabulary, and consistency, and the manual compression effort is not justified when prompt caching provides a more effective and reversible cost reduction.