What are the best models for long-context tasks?

Currently, Gemini 3 Pro supports the largest window (1M+ tokens), while Claude 4.5 and GPT-5.2 are noted for their high retrieval accuracy and resistance to degradation at mid-to-high token counts.

How can I prevent context poisoning in my AI agents?

Mitigate poisoning by implementing tool output verification, truncating context when errors are detected, and using explicit 'correction' prompts to reset the model's internal state.

Does a larger context window always improve agent performance?

No. Larger contexts can introduce 'context distraction' where irrelevant information competes for the model's attention budget, often leading to higher error rates and increased latency.

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

It is a documented behavior where LLMs recall information at the beginning and end of a context window with high accuracy, but struggle to attend to information buried in the center.

Context Degradation Diagnostics

Name: Context Degradation Diagnostics
Author: shipshitdev

byshipshitdev

•

数据科学与机器学习

Identifies and mitigates performance loss in large-context agent systems using proven architectural patterns and empirical benchmarks.

This skill provides specialized knowledge for managing the inherent limitations of long-context windows in LLMs, specifically addressing failure modes like 'lost-in-middle,' context poisoning, and distraction. It enables developers to diagnose why agent performance drops during extended conversations, provides specific mitigation strategies such as context compaction and isolation, and references the latest empirical benchmarks for models like Claude 4.5 and GPT-5.2. It is an essential tool for anyone building complex RAG pipelines or long-running autonomous agent systems where context reliability is critical.

主要功能

01Diagnoses 'lost-in-middle' attention decay patterns

02Detects context poisoning and recursive hallucination loops

03Provides mitigation strategies like compaction, masking, and partitioning

04Analyzes context clash and distraction in multi-source retrieval systems

05Includes model-specific degradation thresholds for GPT-5.2, Claude 4.5, and Gemini 3

0610 GitHub stars

使用场景

01Debugging hallucination spikes in long-running autonomous agent sessions

02Optimizing RAG pipelines to ensure high-accuracy retrieval from large datasets

03Architecting resilient multi-agent systems that require shared context management

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add shipshitdev/library context-degradation

For use in Claude.ai and ChatGPT

Download Skill