I started Skills Orchestrator because I kept rediscovering skills I already had installed. I wanted to type what I was working on and get a few local files worth reading, along with the words that made them match.
The tricky part was deciding when a match was too weak to show. On the current exploratory set, the hook stayed quiet on all four prompts that called for silence, but found an expected skill for only seven of fourteen prompts that needed one. Both sets are in the repo, so you can reproduce the results and add cases from your own setup.
If you use agent skills in a language other than English, I’d especially like to see a short prompt where the expected skill seems obvious to you. That’s a good place to start contributing.
Top comments (1)
The number I would keep on the front page is 7 of 14, not the 4 of 4. Logging the prompts that needed a match and got silence is a recall measurement, and most catalogues only publish the precision side, because misses are invisible by construction: nothing happens, so nothing gets written down. I run an AI tool directory — 4,323 tools, 92 categories, 7,006 category links, 2,100 tools sitting in more than one category, averaging 1.62 each — and that multiplicity is exactly why my "search returned something" checks looked healthy for months. A weak match still returns a row. Meanwhile three of those categories turned out to hold zero tools and were still in my sitemap, which is the retrieval equivalent of your silence set: the failure mode is a confident empty answer nobody measured. The part of your design I am copying is storing the words that made a match rather than only the match. An explanation is the only artefact that lets a reader falsify a retrieval they did not choose, and it is also the only thing that turns your 7-of-14 into a test suite instead of a anecdote.