This is a diagnostic framework for when your agent starts acting weird during long conversations. It breaks down context failures into five specific patterns (lost-in-middle, poisoning, distraction, confusion, and clash) rather than treating everything as random breakage. The core insight is treating the attention U-curve as a design constraint: stuff in the middle of your context window gets 10-40% less attention, so you architect around it. Most useful when you're debugging production agents that work fine for three turns then fall apart, or when you're trying to figure out why retrieved facts aren't being used even though they're clearly in context. Comes from muratcankoylan's agent skills collection, adapted by flora131.
npx skills add https://github.com/flora131/atomic --skill context-degradation