Version 1.0.0. Free, and MIT licensed. From Victor Collective.
Agent skills rot. You write ten or twenty of them, a few never fire, a few fire on everything, and some quietly state things that stopped being true months ago. This kit is a small, closed loop for keeping a skill corpus sharp: one standard, one habit, one audit.
It is written for Claude Code, and it should work with any tool that reads SKILL.md skills.
Leave your email and the zip is yours. We keep the address, the time, and which kit. We do not send mail about this.
skill-design The standard
How to write a description that triggers, what belongs in the frontmatter, how big a skill should be, and the four ways a corpus rots. Use it whenever you write or edit a skill.
skill-harvest The habit
After work ships, it gathers what you learned and routes each item to a skill, an enforcement check, or the bin. You get a numbered list. Nothing is written until you accept it.
skill-curator The audit
A read-only agent that runs five passes over the whole corpus (trigger quality, drift, size, overlap, staleness) and ends with a prune list. Run it on a cadence, say once a month.
install.sh The install
Puts the skills and the agent where your tool looks for them. It will not overwrite a real file, and running it again is safe.
build-manifest.ts The evidence
Builds a manifest for any skill corpus: version, line count, description length, warnings, and git history when there is any. The curator reads it.
Unzip it, then run ./scripts/install.sh from the folder. Bash is enough for the install. The manifest builder needs Node 22.6 or newer (or Bun). If your tool has no subagents, paste the body of agents/skill-curator.md into a prompt and run it by hand.
The kit will not write or delete anything in your corpus on its own. It does not phone home, and it does not replace tests. If a rule must never slip, enforce it with a lint rule, a hook, or a test.
If you want the same kind of review for the agents themselves, Victor Collective runs an agent-reliability audit.