context-degradation

AI context degradation occurs when models lose accuracy by ignoring information in the middle of large contexts, or suffer from context poisoning, distraction, confusion, and clash. This skill diagnoses these specific failure patterns to prevent irrelevant responses.

To fix the lost-in-middle phenomenon, you can apply architectural patterns and mitigation strategies like reordering information or summarizing key points. This skill provides practical guidance to maintain performance within large context windows.

You analyze LLM context for performance degradation by evaluating the conversation against empirical benchmarks and known failure patterns. This skill examines the context to pinpoint causes like distraction or clash and suggests targeted fixes.

The four-bucket approach is a practical strategy provided by this skill to manage large contexts effectively. It categorizes information to mitigate context poisoning and prevent AI failures caused by overwhelming context windows.

You should use context mitigation strategies when an AI agent starts producing irrelevant or incorrect responses during a long conversation. This skill helps identify when context limits are exceeded and applies architectural patterns to resolve it.