prompt-caching

This skill provides strategies for caching LLM prompts and responses to reduce costs and improve performance. It covers prompt prefix caching (like Anthropic's native caching), full response caching, and Cache Augmented Generation (CAG). It's designed for situations where prompts have repeated prefixes, responses are likely to be similar, and semantic similarity matters more than exact match.

LLM caching differs from traditional caching. Prompts have prefixes that can be effectively cached, responses can vary based on parameters like temperature, and semantic similarity often plays a crucial role in determining cache hits. Effective caching requires careful consideration of the appropriate caching level (prefix, response, or both) and robust cache invalidation strategies.

This skill helps you avoid common pitfalls like caching with high temperature settings, neglecting cache invalidation, and indiscriminately caching everything, ensuring optimal cache utilization and accuracy.

What It Does

Implements caching strategies for LLM prompts and responses, including prefix caching, full response caching, and Cache Augmented Generation (CAG).

When To Use

When you need to reduce LLM costs, improve response times for similar queries, or leverage pre-cached documents instead of real-time RAG retrieval.

Installation

Copy SKILL.md to your skills directory