You have hundreds of agent skills installed. Your agent uses about six of them.

Skills are matched to a request by their description. That works at twenty skills. At a thousand it stops working, and the pack you installed last week never fires again.

Stars measure popularity. They do not tell you whether the skills load.

Every pack is cloned and measured on things that can be checked: does the frontmatter parse, is the description specific enough to route on, does it say when to use it, do names match directories, are there duplicates, how much context does an average skill cost. Same commit in, same score out. Nothing here asks a model.

A pack is listed only at 85 or above. Anything below stays listed only with a written exception, and validation fails without one. Scores regenerate every day, so a pack that rots falls off the bar on its own rather than coasting on its stars.

Scores and stars generated 1 mo ago, refreshed daily by workflow · reproduce any row with skillister audit <owner/repo>

A long name beats a precise one by accident. The prior fixes that.

Naive matching counts how many query words appear in a skill's name and description. A five-word name contains more of your query than a two-word one without being a better answer, so a thousand-skill pack wins every search it enters. Skillister weights how much of the name your query actually covers, then down-weights hits drawn from very large collections.

$ skillister compare open source a repo

  naive (substring count)                  skillister (coverage + prior)
  --------------------------------------   --------------------------------------
  collecting-open-source-intelligence      open-sourcing
  performing-open-source-intelligence...   collecting-open-source-intelligence
  open-sourcing                            performing-open-source-intelligence...

  naive picked from "Anthropic-Cybersecurity-Skills" (753 skills), skillister picked from "open-sourcing" (1 skill).

The gain shows up on queries built from common words, which is exactly where a large pack's keyword-dense names crowd out a short precise one. On queries containing a distinctive term the two agree, because there was no ambiguity to resolve. Run skillister find --naive against your own machine and judge it yourself.

19 routes, 60 prompt evals, zero routes that point at nothing.

A route maps a task shape to exact skill names. Each declares prompts that must fire it and prompts that must not, and CI runs all of them. It also asserts every routed name resolves in a real built index, so a route can never recommend a skill nobody installed.

Curation means saying no in public.

Works with whatever you already run.

$ npm install -g skillister
$ skillister index
indexed 1203 skills from 6 roots

$ skillister route "the login page is broken"
- "systematic-debugging" + "diagnosing-bugs" :: Hypothesis loop, no
  guess-patching. Find the root cause before editing.

$ skillister init claude-code   # hooks
$ skillister init generic       # AGENTS.md snippet

Claude Code gets hooks that arm the index at session start and route every prompt. Codex, Cursor, Gemini CLI and opencode get an instruction snippet. One index serves every host CLI on the machine at once.