What problem does it solve? Long-running agent conversations silently degrade: instructions get lost in the middle of the context, hallucinated content poisons later decisions, and contradictory facts produce inconsistent outputs. This Skill gives you a structured framework to identify which failure pattern is occurring and apply the right mitigation. ## Core Features & Use Cases - Five-Pattern Diagnosis: Classify failures as lost-in-middle, context poisoning, distraction, confusion, or clash, each with distinct symptoms and recovery steps. - Four-Bucket Mitigation: Apply Write (external scratchpads), Select (retrieval filtering), Compress (summarization), and Isolate (sub-agent separation) strategies to reduce context load. - Degradation Thresholds: Reference token-count thresholds per model to know when to trigger compaction before quality collapses. - Use Case: An agent that performed well early in a session starts contradicting earlier decisions and misusing tools. Use this Skill to trace the failure to context poisoning, cut the context back to a verified checkpoint, and restart with only validated information. ## Quick Start Ask the agent to diagnose why output quality dropped in this long conversation and recommend context mitigation steps.
