Context Degradation

If you've ever watched an agent slowly lose the plot over a long conversation, this explains why. It covers five degradation patterns: lost-in-middle (that U-shaped attention curve where middle content gets ignored), context poisoning (errors that compound through repeated reference), distraction (irrelevant info eating your attention budget), confusion (model can't tell which context applies), and clash (contradictory information derailing reasoning). The practical bit is knowing these are predictable, not random failures. When your agent starts hallucinating after 50 messages or calling the wrong tools, you're probably hitting one of these patterns. Useful for debugging production issues and understanding why careful context management actually matters at scale.

skills.sh installs: 290
GitHub stars: 39.4K
last updated: Jun 3, 2026

View on GitHub

Install

$ npx -y skills add sickn33/antigravity-awesome-skills --skill context-degradation --agent claude-code

Installs into .claude/skills of the current project.

Files

Context Degradation Patterns

Language models exhibit predictable degradation patterns as context length increases. Understanding these patterns is essential for diagnosing failures and designing resilient systems. Context degradation is not a binary state but a continuum of performance degradation that manifests in several distinct ways.