Skills That Manage Skills: The Meta Layer Arrives
An ecosystem is real when it grows tools for itself. Agent skills passed that threshold a while ago with skill creators — and now a second wave is arriving: skills that audit, index, and route other skills at runtime.
The three tools
skill-link-check — 9.1 — audits the skill installations on a machine: it walks project and global .claude/skills layouts, verifies that references resolve and linked resources exist, and reports the broken links and orphaned files that accumulate the moment you install more than a handful of skills. skillsmp — 9.1 — is an index layer: it searches a 1.6M-plus corpus of public SKILL.md files across Claude Code, Codex, and ChatGPT ecosystems, filtered by keyword, category, and job. using-superpowers — 8.9 — operates at conversation start: it establishes how the agent should find and use skills for the rest of the session, forcing skill discovery before improvisation.
They form a stack of their own. skillsmp answers "does a skill for this exist?" using-superpowers answers "which skill should I be using right now?" And skill-link-check answers "is my local collection actually healthy?" — plus skill-creator (9.1) upstream for making new ones. Nothing in the stack produces content, code, or analysis. It is infrastructure for the skill ecosystem itself.
The scores: strong where it matters
The dimension lines are revealing. skill-link-check posts workflow 10 and content 10 — a perfect pair — with security 10 and engineering 9. An audit is the ideal skill shape: deterministic, ordered, with an obvious completion criterion and an obvious output. skillsmp is close behind at content 10, workflow 9.6, engineering 9. using-superpowers carries the family's highest structure score at 10 but the lowest workflow at 7.6 — routing is judgment, not procedure, and the dimension correctly notices.
The shared soft spot is trigger quality: 7.0, 7.0, 8.0. That number encodes a genuine structural problem rather than sloppy writing. A skill-creation trigger is easy to write because the user says "make me a skill." But "audit my skills" competes against the user not knowing an audit is possible, and "search the corpus before you improvise" competes against the agent's default behavior of just improvising. Meta-skills fire only when the user is already thinking in meta terms — the population that most needs link-checking is the population least likely to ask for it. The chicken-and-egg is measurable, and it is measured here at 7.0.
Why the meta layer is forming now
The timing is not arbitrary. Skill collections rot: links break, references dangle, duplicates accumulate. The failure mode scales with adoption, and adoption has scaled. You only need a link checker once you have enough links to check — and the existence of a 1.6M-SKILL.md index says the corpus long since passed the point where any agent, or any human, can hold it in their head. Search infrastructure appears when the searchable mass exists. Audit infrastructure appears when the installed base decays fast enough to notice.
The same pattern ran in every earlier ecosystem: package managers grew npm audit, pip check, and dependency doctors only after the registries grew large enough to rot. The agent-skill ecosystem compressed that timeline into months.
The uncomfortable implication
A healthy meta layer is also a load-bearing dependency. When using-superpowers decides which skills get loaded, it becomes the router for everything else — an 8.9-scored component sitting in front of 9.5-scored skills. That is fine while the routing is simple, and it will not stay simple. The meta layer will need the same evaluation discipline as the layer it manages, or the ecosystem will inherit its blind spots wholesale. The tools that manage the tools deserve the most scrutiny of anything in the directory — precisely because they are the tools least likely to be scrutinized.