Skills Janitor

The problem Skills Janitor addresses is real and easy to underestimate: every installed skill's description loads into context on every request whether or not you use it, and every configured MCP server loads its full tool schema the same way. A collection built up over a year of "this looks useful" installs can quietly tax every prompt you send. Most people know this in the abstract and never do anything about it, which is exactly the gap this tool targets. What it gets right: the data is measured, not guessed. Usage counts are pulled from real session transcripts and include skills Claude triggered on its own, so a skill you never manually invoked but which auto-fires twice a day does not get flagged as dead. Cross-scope duplicate detection uses Jaccard similarity above a 30% threshold, which catches the common pattern of the same skill existing in both your user scope and a project scope. The v1.7 MCP work is the sharpest part — cross-referencing configured servers against mcp__server__tool records in transcripts turns an invisible cost into a number you can act on, and the eight-week silence window for "connected but never called" is a sensible line. Deletion is staged with a timestamped .bak backup for anything it touches, which is the right amount of paranoia for a tool whose whole job is deleting files. The swipe TUI is the reason anyone will actually use it. Sorting the deck heaviest-and-least-used first means the highest-value deletion is card one, and most people hit their useful pruning in the first 5-10 cards. The honest limitation is that this becomes a self-limiting habit rather than a full audit — you are not going through 178 skills one at a time. The security scan (v1.6+) is a genuine differentiator because a skill is text your agent trusts unconditionally. Checking for instruction-override phrasing, "don't tell the user" directives, instructions buried in HTML comments or zero-width unicode, base64 payloads, and scripts that pipe the network into a shell or read credential stores covers the known bad shapes. The PASS/REVIEW/RISK verdicts are explicitly heuristics, and the tool says so — a RISK means read it before trusting it, not that it is malware. Being calibrated for low noise matters more than breadth here. Where it does not fit: it is Claude Code and Codex only. Cursor, GitHub Copilot, and other assistants get nothing. It is a terminal tool end to end — six slash commands and a TUI — with no web UI and no cloud dashboard for central team management, so a team that wants one shared pane of glass for skill governance will be disappointed. Plugin-installed skills and plugin-bundled MCP servers are flagged for plugin review rather than individually deletable, which is a sensible correctness decision but does mean you cannot fully prune from inside the janitor. And the MCP reporting deliberately stops at provable call counts and last-used dates — no per-server dollar estimates — so if you want a budget line item, you will