What problem does it solve? Teams waste tokens and create inconsistency when every session repeats conventions in prompts or when a single monolithic CLAUDE.md mixes personal preferences with shared rules. This Skill structures persistent agent memory into three cascading levels so each convention lives in exactly one authoritative place. ## Core Features & Use Cases - Three-level hierarchy: Separates memory into ~/.claude/CLAUDE.md (user), <repo>/CLAUDE.md (team), and <repo>/<subpath>/CLAUDE.md (module), with precedence subpath > repo > user. - @imports modularization: Keeps the main file short and cache-friendly by importing small files from docs/ instead of inlining thousands of lines. - Compaction boundary: Distinguishes transient Claude API compaction blocks (transcript continuity) from durable memory, requiring governed review before promoting content. - Use Case: A repo has a 2000-line CLAUDE.md mixing style guides, testing conventions, and personal preferences like "terse commits". This Skill refactors it into a short team file with @imports and moves personal preferences to the home file. ## Quick Start Audit my project's CLAUDE.md files and reorganize them into user, team, and module levels using @imports for the long sections.