How can I prevent context distraction?

Mitigation includes applying relevance filtering, using namespacing to organize data, and moving non-essential information into tool-accessible storage rather than direct context.

Which models does this skill provide benchmarks for?

The skill includes specific degradation thresholds and behavior patterns for Claude 4.5 series, GPT-5.2, and Gemini 3 Pro/Flash.

How does context poisoning affect AI agents?

Poisoning occurs when errors or hallucinations enter the context and compound through repeated reference, creating feedback loops that reinforce incorrect agent beliefs.

What is the 'lost-in-middle' phenomenon?

It is a documented pattern where LLMs show significantly reduced recall accuracy for information located in the center of a context window compared to the beginning or end.

Context Degradation & Engineering

Name: Context Degradation & Engineering
Author: goodnight000

bygoodnight000

数据科学与机器学习

Diagnose and mitigate performance loss in agentic systems caused by large, conflicting, or poisoned context windows.

关于

This skill equips Claude with specialized knowledge to identify and resolve patterns of context degradation, such as the 'lost-in-middle' phenomenon, context poisoning, and information distraction. It provides empirical benchmarks and architectural patterns—including compaction, masking, and partitioning—to ensure agents maintain high reasoning quality during long-running tasks. By applying these techniques, developers can design more resilient agentic systems that handle large context windows without sacrificing accuracy or reliability.

主要功能

Identify and correct 'lost-in-middle' attention patterns
Apply mitigation strategies for context distraction and confusion
Detect and recover from context poisoning and compounding errors
Implement architectural patterns like context partitioning and isolation
Reference model-specific degradation benchmarks for Claude, GPT, and Gemini
0 GitHub stars

使用场景

Debugging agent failures in long-running or large-scale development tasks
Designing production-grade context engineering pipelines for AI agents
Optimizing context placement to improve information recall and reasoning

关于

主要功能

Identify and correct 'lost-in-middle' attention patterns
Apply mitigation strategies for context distraction and confusion
Detect and recover from context poisoning and compounding errors
Implement architectural patterns like context partitioning and isolation
Reference model-specific degradation benchmarks for Claude, GPT, and Gemini
0 GitHub stars

使用场景

Debugging agent failures in long-running or large-scale development tasks
Designing production-grade context engineering pipelines for AI agents
Optimizing context placement to improve information recall and reasoning