Building a cleanup tool for AI agent skills
Skills Janitor audits every skill, subagent and MCP server installed in Claude Code or Codex, shows what each one costs in context, and deletes the dead weight. I launched it in the middle of the skills boom, when almost nobody had built one, and it finished #4 of the day on Product Hunt with no paid promotion.
- ROLE
- Everything: idea, build, launch, maintenance. Code with Claude Code.
- SINCE
- March 2026
- STACK
- bash + python3, zero dependencies, Claude Code plugin, Codex
- STATUS
- Live, v1.8, maintenance mode
Product Hunt and GitHub, as of 2 Nov 2026.
01 THE PROBLEM
35 skills installed, 4 actually used
In early 2026 skills exploded: small packs of instructions you install into Claude Code so the agent knows how to do something. Everyone was installing them, and every installed skill sits in the agent’s context whether you use it or not. I had 35. I used 4.
There were plenty of skills, and almost nothing to clean them up. So I built the janitor: find what is installed, what is duplicated, what is broken, what you never call, and remove it safely.
02 THE LAUNCH
Scheduled a week ahead, then forgotten, then #4
I put the launch on Product Hunt a week in advance and forgot about it. On 14 April it went live on its own, in the middle of the skills boom, and climbed to #4 of the day with 175 upvotes. Product Hunt picked it for its daily newsletter the same day. No ads, no upvote requests, no launch squad.
The comments were the real prize. That evening I wrote down what people asked for, and the next day shipped it: a check for overlaps before you install a new skill, nine commands folded into seven, and support for OpenAI’s Codex next to Claude Code.
| 14 Apr (launch) | 30 |
| 20 Apr | 76 |
| 26 May | 98 |
| 2 Nov | 124 |
03 STAYING HONEST
The tool was blind to half my skills, so I said so
Six weeks after launch I found out it had only ever looked in one folder. Every skill installed as a plugin was invisible to it. Version 1.3 fixed that, and on the same machine the count went from 35 to 157. I posted it as exactly that: “I shipped a tool that was blind to half my skills.”
The first draft of that announcement used a scary percent-of-context number. It was misleading, so I cut it and kept the plain count. A tool that tells you what to delete can’t afford to exaggerate.
04 WHERE IT IS
Finished enough to leave alone
A dozen releases later it covers skills, subagents and MCP servers, reads real usage from session transcripts, scans skills for prompt injection, and backs up everything before it deletes. Other developers send fixes and I merge them. Since version 1.8 at the end of September it is in maintenance mode: it does its job, and it is listed where people look for tools like it.
What I took from it: timing matters more than a launch plan. Being early in a boom, with something small that works, did more than any campaign could have.
05 TRY IT