Any AI chat can now run skill search — and you approve every request

Insight · Mike Arbuzov · SkillFed Research · August 03, 2026

AI-assisted notes · reviewed by SkillFed

No install, no account, no connector. Your chat writes an abstract wish, you paste the link back, and it reads five security-swept skills. The whole request is a URL in plain English — the privacy boundary is something you check, not something you're asked to trust.

Any AI chat can now run wish-based skill search — ChatGPT, Claude, anything with web access. No install, no account, no connector. Your chat works out what the ideal skill for your task would do, sends that as an abstract wish — your code and your plan never leave the window — and gets five matched skills back to read and apply. Every body it reads is a security-swept snapshot we hold and version, not a raw pull from a stranger's repository. And you approve each search before it runs.

Why skills, when the model already knows

A model's knowledge is an average of what was written before its training cutoff. That is fine for things that hold still. It is quietly wrong for everything that moves — framework migrations, packaging defaults, security practice, the tooling flags that changed two minor versions ago. What comes back is a confident summary of a past consensus, delivered in the same tone as a current one. There is no tell.

A skill is the opposite kind of artifact: a dated document that someone maintains, about how one thing is done now. Not a rule to obey, not a script to run — field notes. Most of the value is in a model reading one, which is why consulting is the ordinary case and installing is the exception.

We watched this happen while testing. A model gave solid, confident advice on a Swift 6 concurrency migration. It then read two maintained skills on the subject and revised its own recommendation — the guidance it had given was right for an earlier version and had been overtaken. It also flagged a broken code example inside one of those skills, and a place where the skill contradicted its own reference page. Both of those are the system working as intended: read more than one, and cross-check.

What actually leaves your window

This is the part worth being concrete about, because "we don't send your code" is a thing everyone says.

Your chat does not send your code, your files, or your plan. It sends a description of the capability it is looking for. And on this route, that description is the URL — the whole request, in plain English, in front of you:

https://skillfed.io/api/q/migrate-an-ios-app-to-swift-6-strict-concurrency-and-fix-data-race-warnings

That is everything we receive. No repository name, no file tree, no diff, no prompt history. You can read it before you send it, which is a different thing from being told it is fine.

The discipline is to keep the wish at the level of what skill should exist, never what you are working on:

leaks your work	describes the need
`fix-the-race-in-PaymentQueue-swift-line-88`	`swift-6-strict-concurrency-data-race-migration`
`why-is-acme-checkout-build-slow`	`speed-up-a-javascript-monorepo-build-with-caching`

If a wish cannot be written without quoting your work, it needs abstracting further.

The paste is the gate

Your chat cannot call us on its own. It composes the search link, shows it to you, and stops. Nothing happens until you paste it back.

That looks like friction and it is the whole design. A chat that can reach any URL by itself is a chat that can be talked into reaching one by a web page it happened to read. Making the link come from you means every request is one you saw, in full, before it left. The privacy boundary is not a policy we are asking you to trust — it is a URL you read and approve.

What comes back

Five skills, ranked, each with its license, publisher and update date. Your chat reads the ones that look relevant and applies what fits.

The bodies it reads are snapshots we hold: pinned to a recorded version and swept with automated security tooling before they are served. That is meaningfully better than pulling a file straight from a repository nobody has looked at. It is not a guarantee, and it does not replace your judgement about whether to act on what a skill says.

Two things follow from the word snapshot: you get the exact bytes we indexed and described, and you may not get whatever sits at the repository's HEAD right now. Every result carries its version and date, so you can tell which.

What this does not promise

Skills are third-party open-source files. We index, describe, rank and sweep them — we do not vouch for the people who wrote them, and a listing is not an endorsement. They disagree with each other, sometimes usefully: two independently written skills saying the same thing is current practice; one asserting it alone is a hypothesis worth checking.

The catalog is also uneven. Ask it something outside what it covers and it will still return its closest matches. Your chat should say so when the honest answer is "nothing here fits" — and so should you, if it does not.

Try it

Open any chat with web access and ask it to find you a skill. It will write the link; you paste it back. Or start from a worked one:

https://skillfed.io/api/q/write-good-git-commit-messages

The full protocol, and what the security sweep does and does not cover, is on using SkillFed from a chat. If you work in a coding agent instead — Claude Code, Cursor and the like — it can call the same index directly without the paste step, and install a skill when one is worth keeping. What that index is built on: the census of 60,611 skills in the wild, and the research map behind it.