Reins Audit Skill

Reins Audit turns a settings-file cleanup chore into a conversation. Run /reins-audit, or just ask “audit my Claude Code permissions,” and it walks every standing grant it finds, tells you what each one actually permits, and only touches anything after you say so.

What it does

The skill reads Claude Code’s settings sources, classifies each standing permission grant into one of 12 risk categories, and groups the results by verdict: remove, review, or keep. Anything flagged gets a one-sentence blast-radius narrative — what a hijacked session could do with that grant, not just a risk label. If you ask for cleanup, it lists exactly which rules would be removed and waits for explicit confirmation before changing anything.

How it works

Two modes, same taxonomy. Standalone, the skill classifies by reading its own bundled reference files (references/taxonomy.md, references/remediation.md, references/policy-baseline.md) and applying that logic by hand against the settings files it reads directly — no MCP server required. When the Reins MCP server is connected, the skill hands classification to it instead: the server’s deterministic regex engine, secret redaction, and blast-radius scoring replace the manual pass, and confirmed revocations go through the server’s reins_revoke tool, which re-verifies each rule is still present, backs up the file, and writes atomically. Either way, the skill never displays a secret’s actual value and never modifies read-only sources (the Claude state file or managed settings) — it can only report what’s there.