Extended Thinking

Anthropic markets it as a reasoning feature. I use it as a debugger. When Claude misnames a token or mislabels a component, the scratchpad tells you exactly why.