context-degradation

AI context degradation occurs when language models perform poorly as processed information increases, manifesting as lost-in-middle, context poisoning, distraction, confusion, or clash. This skill identifies these specific patterns to explain why your agent fails to retrieve early conversation details.

Context degradation thresholds are approximate token limits where language models begin losing information. This skill provides specific degradation thresholds for various LLMs to help you anticipate when large contexts will cause agent failures.

Mitigation strategies for context poisoning involve restructuring prompts and applying best practices for managing large contexts. This skill offers techniques to prevent or recover from degradation by optimizing how information is presented to attention mechanisms.

The lost-in-middle problem is a context degradation pattern where language models fail to use information positioned in the middle of large contexts. This skill explains how attention mechanisms cause this issue and provides best practices to manage it.

You should restructure prompt data when context approaches degradation thresholds or when debugging agent failures related to attention mechanisms. This skill helps identify when large contexts cause confusion or clash patterns and suggests ways to improve performance.