Claude

Your team's operating knowledge (how you name tokens, review work, and decide what good is) lives in people's heads. Put it in a repo of skills and playbooks and your least AI-fluent teammate works like your best one, while agents execute against the same source. This guide builds it, with the maturity ladder and review model that keep it honest.

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.

How I label 20 component names by hand and get Claude to pattern-match the rest. Practical setup for component naming, token naming, and any classification task.

A 5-version walkthrough showing how a bare prompt becomes a production-ready design audit. Built on the same iterative method Anthropic teaches internally.