context-degradation

This skill is designed for AI engineers and developers who need to diagnose and fix performance failures in large-context agent systems. Context degradation is rarely a binary failure; it is a continuum of predictable patterns that undermine LLM reasoning, retrieval-augmented generation (RAG), and multi-step agentic workflows. By analyzing attention mechanics—specifically the U-curve effect where middle tokens suffer from reduced recall—this skill provides actionable strategies to restructure information placement, filter irrelevant noise, and isolate conflicting data streams. It is an essential tool for debugging unpredictable model behavior, incorrect tool calls, or hallucinations that persist across long conversation turns.

• Detection and mitigation of 'lost-in-middle' phenomena by optimizing information positioning within the context window.

• Prevention of context poisoning by establishing rigorous provenance tracking and truncation protocols for tool outputs and retrieved documents.

• Aggressive relevance filtering to address the step-function performance degradation caused by distractor documents.

• Task isolation and context segmentation to prevent 'context confusion' and cross-contamination of task-specific constraints.

• Resolution strategies for context clash, including priority rules, source precedence, and contradiction filtering.

• Use this tool when agents experience unexpected output degradation or exhibit 'forgetfulness' during complex, multi-turn reasoning sessions.

• For best results, integrate this diagnostic framework into your evaluation pipeline to monitor claim provenance and tool alignment.

• Inputs typically include raw conversation history, retrieved context windows, or task logs; outputs include structural recommendations for prompt engineering, document placement, and system-level context resets.

• Note that this skill treats context as an engineering constraint rather than a model bug, emphasizing the management of attention budgets through structural markers, explicit headers, and efficient tool-call architectures.

• Suitable for systems utilizing long-context models where attention-sink tokens and memory scarcity limit performance in production-grade deployments.