links: repo not public yet | no live url yet
the first real test bed for the updated blueprint is an ai skills directory.
the idea is simple: skills are becoming shareable software, but the trust model is terrible. a skill can look like a helpful workflow and still carry risky instructions, hidden dependencies, or assumptions that do not belong in someone else’s environment.
so the directory cannot just be a pile of downloads. the interesting part is the review surface.
i scaffolded the app as an Astro workspace with React islands for the places that need real interaction. that feels like the right split. most of the site is directory and content. upload, validation, and review flows can be islands instead of turning the whole thing into a client app.
the first version should probably accept both uploaded skill folders and git repo links. uploads are better for casual users. repo links are better for maintainers who already version their work.
the part i care about most is the validation language. i do not want a green checkmark that pretends to mean “safe.” i want clear signals: schema passes, metadata present, suspicious instructions found, network or dependency warnings, manual review needed.
the directory is useful only if it teaches people not to blindly trust the thing they just downloaded. discovery is half the product. skepticism is the other half.