Every team is now writing, sharing, and collecting Claude Code skills — markdown files that define how your agents think and act. As that library grows, so do the problems: duplicated skills, conflicting instructions, accidental PII, and more. TidyRepo audits your entire skill repo automatically, so your agents always get the best version of every skill.
Findings will appear here…
At a few dozen files, a knowledge base works fine. Markdown is readable, authors know each other, and updates stay fresh. Then the org scales. At a few hundred files (let alone a thousand), the same six problems appear in every team, every domain, every company.
The result: people stop trusting the knowledge base. They search, find three different answers, and ask a colleague instead. The whole point collapses.
TidyRepo catches issues humans miss at scale, not because humans are careless, but because no human can hold a thousand files and all their contradictions in their head at once.
TidyRepo is a single Claude Code skill with three focused commands. Run one for a targeted check, or chain them for a complete pass. Each command produces a structured report: findings grouped by severity, not a flat list of noise.
/tidyrepo piiAvailableScans every file for email addresses, phone numbers, API keys, internal credentials, and named personal identifiers. Two-layer detection: regex for structured patterns, NER model for names and addresses.
Essential for any repo with external contributors or public access.
/tidyrepo dupesAvailableGroups semantically similar files using embeddings before comparison, so expensive analysis runs only where overlap is likely, not across every pair. Flags files for consolidation with similarity scores.
High value during taxonomy cleanup and repo consolidation.
/tidyrepo conflictsAvailableSurfaces conflicting guidance across documents and groups conflicts into resolution buckets. Each finding includes both sides of the conflict and a suggested resolution path, not just a flag.
Critical before publishing or sharing the knowledge base externally.
More commands on the roadmap: staleness detection, coverage gap analysis, ownership inference, and more.
Every finding includes the file, line number, what was detected, and a concrete remediation step. Sorted by severity so the most urgent issues are always first.
A naive audit (load all files, compare everything against everything) breaks down fast. At a few hundred files you're already exceeding context windows. At a thousand, brute-force comparison is computationally intractable. TidyRepo assumes the repo is large and designs around that constraint.
Related files are grouped by semantic similarity before any deep analysis. Expensive comparisons (contradiction checks, duplication scoring) run only inside clusters where issues are likely to exist.
Within each cluster, targeted checks run: contradiction detection, duplicate scoring, and PII scanning. Each check is focused and bounded, not a full-corpus scan.
Findings are grouped by severity and theme, not flattened into a wall of issues. Where possible, likely ownership is attached so work can be assigned, not just discovered.
The output is designed to be opened on a Monday morning and acted on the same day. Not filed and forgotten.
TidyRepo runs inside Claude Code. You need Claude Code installed and authenticated. No external server, no separate API key setup beyond your existing Claude auth.
Install Claude Code if you haven't already
npm install -g @anthropic-ai/claude-codeClone TidyRepo once you have been added to the repo
git clone https://github.com/rachhek/tidyrepoOpen Claude Code in your KB repo, pointing to the plugin
claude --plugin-dir ~/tidyrepoRun your first audit
/tidyrepo piiTidyRepo is currently in private early access. Leave your email and we'll add you to the GitHub repo — ready to clone and run in minutes.
Submit your email in the early access section above. We'll add you to the private GitHub repo, after which you can clone it and follow the install steps to be up and running in minutes.
Any markdown-based knowledge base. Works on plain GitHub repos, Obsidian vaults, Notion exports, GitBook sources, internal wiki exports, and Claude Code skill repos. If it's markdown files in a directory, TidyRepo can audit it.
Your repo content is processed by Claude (Anthropic's API), subject to Anthropic's standard data handling and privacy terms. For enterprise accounts with zero-retention agreements, content is not used for training. No third-party services or external storage are involved beyond Anthropic's API.
TidyRepo is designed around the assumption that the repo is large. Semantic clustering groups related files before any expensive analysis runs, so the system stays tractable at hundreds or thousands of files. It does not attempt to load everything into a single context window.
Each skill produces a structured report: findings ranked by severity (critical, high, medium), grouped by category, with the specific files involved and a suggested remediation path. Where possible, likely ownership is inferred from git history. The goal is a report you can act on immediately.
Yes. Each skill is a standalone slash command. Run /tidyrepo pii on its own if PII is your only concern today, or run all three in sequence for a full audit pass.