Context Degradation Diagnostics

关于

The Context Degradation skill provides Claude with a specialized framework for identifying and fixing reliability issues in long-context AI systems. It offers deep insights into the 'lost-in-middle' phenomenon, context poisoning from hallucinations, and the 'distractor effect' where irrelevant information degrades reasoning. By applying this skill, developers can implement architectural patterns like context partitioning, isolation, and compaction to maintain agent performance even as context windows approach their limits.

主要功能

• Diagnoses context poisoning and provides recovery paths for compounding errors.
• Identifies 'lost-in-middle' patterns to optimize critical information placement.
• Analyzes context clash between contradictory multi-source data points.
• Provides model-specific degradation benchmarks for Claude, GPT, and Gemini.
• 0 GitHub stars
• Recommends mitigation strategies like relevance filtering and task segmentation.

使用场景

• Optimizing RAG systems by identifying and removing context distractors.
• Debugging agent failures where responses become irrelevant during long conversations.
• Designing resilient agent architectures that handle 100K+ token windows reliably.