context-degradation

This skill provides a structured engineering framework for identifying and resolving performance degradation in AI agent systems. Context engineering is the discipline of optimizing the information landscape within a model's context window. This skill is specifically designed for developers and AI engineers who observe unpredictable agent behaviors, such as hallucinations, tool misuse, or failure to follow instructions in long-running sessions. By treating context as a finite, load-sensitive resource governed by attention mechanics, it helps practitioners shift from trial-and-error prompting to rigorous, systemic failure mitigation.

• Detection and diagnosis of the 'lost-in-middle' phenomenon where critical information is ignored if placed in the center of long contexts.

• Strategies to prevent 'context poisoning' by validating tool outputs and retrieved documents before they enter the model's reasoning window.

• Techniques for aggressive relevance filtering and curation to prevent 'context distraction' caused by irrelevant background data.

• Implementation of task isolation patterns to prevent 'context confusion' where constraints from different objectives bleed into one another.

• Methods for resolving 'context clash' when multiple contradictory sources exist within the prompt window.

• Users should activate this skill when investigating agent performance drops, debugging incorrect tool calls, or evaluating system design for production-scale context management.

• The skill expects input regarding the agent's behavior, current context length, and symptoms like contradiction or ignored instructions; it outputs specific architectural adjustments, placement strategies, and filtering rules.

• Practical tips include utilizing the U-curve attention principle by placing core instructions at the very beginning or end of the prompt and using structural anchors like markers or headers.

• Constraints: The skill operates on fundamental LLM attention architecture; therefore, effectiveness is contingent on the specific model's attention span and recall capabilities. It prioritizes a 'curation over accumulation' approach, suggesting that developers move non-essential context behind tool calls rather than flooding the primary window.