What problem does it solve? Copy-pasted guidance across many CLAUDE.md files drifts over time, so duplicated sections become a correctness problem: some copies describe tools that changed, files that were deleted, or targets that never existed. This Skill provides a disciplined measure-verify-converge-lift procedure to deduplicate that content without promoting false claims to ancestor files. ## Core Features & Use Cases - Duplication Measurement: Ships scripts/claudemd_variance.py, which walks a tree (never grep, so gitignored files are found), splits every CLAUDE.md into sections, hashes whitespace-normalised bodies, groups identical variants, and computes each group's true common ancestor. - Verification Before Convergence: Prescribes checking every claim against ground truth (files, commands, skills, mechanisms) before keeping it, and composing canonical text from verified lines rather than picking the most-copied variant. - Safe Lifting Rules: Guards against double headings, per-file reachability checks, and loss of repo-specific rows when lifting a section to a common ancestor. - Use Case: A monorepo has the same workflow section pasted into 20 project CLAUDE.md files with three drifted variants; run the variance script, verify each claim, converge to one correct text, and lift it to the shared ancestor. ## Quick Start Run uv run scripts/claudemd_variance.py --root ~/src --json to measure duplicated CLAUDE.md sections, then follow the verify, converge, and lift steps to consolidate them.
