How does this skill help with agent failures?

It identifies specific failure patterns like context poisoning or distraction that cause agents to produce incorrect outputs as context length increases.

Can this skill fix LLM hallucinations?

It helps detect 'context poisoning,' where errors compound through reference, and suggests strategies like context truncation or explicit re-evaluation to stop hallucination loops.

What is the lost-in-middle phenomenon?

It refers to a U-shaped attention curve where LLMs recall information at the beginning and end of a prompt significantly better than information buried in the middle.

When should I use the context-degradation skill?

Use it when an agent's performance degrades unexpectedly, when debugging irrelevant outputs, or when designing systems that require high reliability with large datasets.

Which models are covered in the degradation benchmarks?

The skill provides diagnostic data and thresholds for major models including Claude 4.5, GPT-5.2, and the Gemini 3 series.

Context Degradation Diagnostics

Name: Context Degradation Diagnostics
Author: muratcankoylan

bymuratcankoylan

•

5,498

数据科学与机器学习

Diagnoses and mitigates performance failures in agentic systems caused by large context windows, attention loss, and information noise.

关于

The Context Degradation skill provides a comprehensive framework for identifying and resolving predictable failures in large-scale context engineering. It equips Claude with the patterns to recognize 'lost-in-middle' phenomena, context poisoning, and distraction, offering architectural strategies like compaction and isolation to maintain high performance in production agent systems. This is essential for developers debugging complex multi-agent workflows or optimizing long-context retrieval accuracy where standard attention mechanics may fail.

主要功能

Detects and remediates context poisoning and hallucination feedback loops
5,498 GitHub stars
Resolves context clashes and contradictory multi-source information
Identifies context distraction from irrelevant information or noise
Diagnoses 'lost-in-middle' attention patterns in long context windows
Provides model-specific performance benchmarks and degradation thresholds

使用场景

Debugging agent performance drops during long-running conversations
Optimizing retrieval-augmented generation (RAG) pipelines for accuracy
Designing resilient context management for multi-agent architectures

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add muratcankoylan/agent-skills-for-context-engineering context-degradation

For use in Claude.ai and ChatGPT

Download Skill

GitHub

关于