CLAUDE.md Improver
Audit and precisely improve CLAUDE.md files across a repository.
The skill explicitly states that it can write CLAUDE.md files and requires a report followed by user approval; its update guidance requires diffs and preservation of structure, providing some consent and review controls. Points are deducted because Bash and Edit are broad capabilities, with no explicit least-path permissions, sensitive-data handling, data-flow boundaries, backup/rollback mechanism, or limits on external side effects.
The workflow is divided into discovery, assessment, reporting, confirmation, and update phases, with checklists for stale commands and other failure conditions. Points are deducted because this is static documentation without a test suite or execution evidence; the find command has unparenthesized operators that may create scope ambiguity; and there are no clear failure messages for empty results, duplicates, more than 50 files, or unverifiable commands.
The name, audience, and trigger semantics are clear, covering root, monorepo, package-level, and nested CLAUDE.md files while limiting updates to targeted changes. Points are deducted because non-fit cases, input-size limits, scoring boundaries, and machine-readable output are unspecified, and Chinese-language workflows are not addressed. The core function does not depend on overseas services, so no mainland-China reachability deduction is applied.
The documentation has clear information architecture with progressive workflow, scoring criteria, templates, update guidance, and a validation checklist; the repository supplies an Apache-2.0 license and the stated official provenance supports attribution. Points are deducted because the skill lacks its own version, changelog, installation/dependency notes, explicit maintenance responsibility, or dedicated update path; it also lacks guidance for missing references, FAQs, and compatibility notes.
The stated goal is to audit and improve CLAUDE.md files, and the workflow includes quality scores, per-file reports, specific issues, recommended additions, diffs, and user confirmation, which could produce directly reviewable changes. Points are deducted because commands, paths, and outcomes are not statically verified; applying updates depends on approval; and no representative completed output is provided, leaving potentially substantial manual review.
The skill supplies an auditable process, rubric, templates, diff format, and validation checklist, while repository CI provides limited maintenance-context evidence. Points are deducted because there is no skill-specific test suite, execution log, or third-party reproduction evidence, and completion of codebase cross-checking is prescribed but not demonstrated; therefore the score remains within the static maximum of 5.
- The skill can modify repository files through Bash and Edit; confirm target paths, review diffs, and retain a recoverable copy before applying changes.
- Commands and file references were not executed or verified in this static review, so quality scores may differ from the repository's actual state.
- Discovery is capped at 50 results and may omit deeper or additional CLAUDE.md files.
- Handling for sensitive data, symlinks, permission errors, and repositories with no CLAUDE.md files is unspecified.
What it does & when to use it
This development skill audits and improves CLAUDE.md files. It scans a repository for CLAUDE.md, .claude.md, and .claude.local.md files, then evaluates them against defined quality criteria. It produces a file-by-file quality report and makes targeted edits only after user approval. It is suited to Claude Code users maintaining project context, workflow guidance, and project memory.
It uses Glob, Grep, Read, and Bash to locate and inspect CLAUDE.md files across project roots, nested directories, package directories, and documented global-default locations. It scores commands and workflows, architecture clarity, non-obvious patterns, conciseness, currency, and actionability, then produces a report with scores, grades, issues, and recommended additions. After the report, it waits for confirmation and uses Edit to apply targeted updates to CLAUDE.md files.
- A monorepo team needs to check whether root-level and package-level CLAUDE.md files describe the current architecture.
- A development team has migrated or refactored a project and needs to identify stale commands, architecture notes, or test instructions.
- A developer taking over an unfamiliar repository wants a quick assessment of whether its project-context documentation is actionable.
- A Claude Code user wants CLAUDE.md maintenance or project memory optimization.
Pros & cons
- Covers CLAUDE.md files across multiple directory levels.
- Provides explicit scores, grades, issues, and improvement recommendations.
- Requires a report and user approval before writing changes.
- Targets concrete content such as commands, architecture, testing, and project-specific rules.
- It is centered on CLAUDE.md and Claude Code workflows, so other clients require adaptation.
- The source provides no test suite, platform compatibility validation, or failure-recovery guidance.
- The assessment refers to references/quality-criteria.md and templates.md, whose contents are not included in the source material.
- The skill can write to CLAUDE.md files, so recommendations and diffs should be reviewed before approval.
How to install
The skill is located at https://github.com/anthropics/claude-plugins-official in plugins/claude-md-management/skills/claude-md-improver/SKILL.md. The README states that plugins can be installed through Claude Code's plugin system with /plugin install {plugin-name}@claude-plugins-official; the source does not provide a standalone installation command or the exact plugin name for this skill.
How to use
In a Claude Code session where the skill is installed and enabled, use a request such as “Audit this repository’s CLAUDE.md files,” “Check and improve CLAUDE.md,” or “Optimize project memory.” The skill presents a quality report first and edits files only after approval.