Agent-ready docs index

A <?catalog?> in CLAUDE.md keeps one summary line per tracked doc, so an agent reads a few thousand tokens of metadata up front and opens only the files a task touches.

AI coding agents pay context tokens for every file they read. Loading a whole docs tree up front wastes most of them.

Progressive disclosure inverts that. A <?catalog?> in the agent file keeps one summary line per tracked doc. The agent skims the index, then opens only the files a task touches. mdsmith check flags the index when it drifts; mdsmith fix rebuilds it.

mdsmith’s own CLAUDE.md is the live example. Its catalog compresses the docs tree to a few thousand tokens of metadata.

See the progressive disclosure guide for the step-by-step recipe.