Dockerhub for Skill.md | Hacker News Reader

Dockerhub for Skill.md

skillregistry.io

This is nothing like Dockerhub and, I'm sorry, but it's seriously useless. In its current state its worse than basically anything else.

You have no versioning, no automated or simplified update, no way to verify the authors, etc. The "installation" is literally just a wget.

This is a really poor solution for the moment, and honestly I think for the forseable future. I don't see how anything beyond git is necessary for skills management.

Most of the skills currently hosted are also really bad. They are just a duplicate of the information that MCP would give the models.

Couple of problems with git.

In the enterprise, RBAC is a royal pain. You give out a URL and it's hard to know if the consumer can fetch it.

URLs are absolute, there is no resolution by name. Compounded further if you want transient dependencies (maybe not needed in this instance though).

In your project, you end up hardcoding the https/ssh scheme.

Homebrew has been using git in the backend to manage its database of package formulas since its inception. No reason it wouldn't work here as well

Homebrew's built a package manager on top of git. I'm talking about platforms that generate built artifacts and have package managers with dependency resolution to fetch them.

mcp will probably be left behind in the future. it was a bad design from the start. anthropic themselves released skills to "fix" the mcp mess. skills are very new but the idea is great. we still are early days but i think it could allow models to use tools more effectively.

we're planning to add an installation step + auth step (which many of the skills require) so that that part get's handled in one single step instead of having to do everything manually

If you want to share skills using something that has versioning, automatic updates, and focused on teams vs the internet at large, consider sx - https://github.com/sleuth-io/sx

I think calling it "official" might be giving users the wrong impression here.

EDIT: It doesn't help that the skills have a checkmark next to the company's name, even though these skills weren't created by the respective companies.

Agreed, I had to retract my upvote for that reason.

I do like the idea of crowd-sourced collections of resources like skills.

It might be more useful if it was an index of skills managed in GitHub. Sort of like GitHub actions which can be browsed in the marketplace[1] but are ultimately just normal git repos.

[1] https://github.com/marketplace?type=actions

i thought of that but i didn't want to build a job to migrate that to the db. maybe we'll go that route.

Have we finally tricked devs and companies into writing good documentation by making it into an AI thing?

This has got to be the dream scenario for technical writers and historians who have a hard time getting the business to invest into their work. Better writing and comprehensive documentation make all your devs using AI write better code as well as easier adoption by your customers.

All my skills are AI-written.

AI agent skills are very useful. Unlike MCP they do not waste context. Most of the time I am building skills that are very particular to my project. But occasionally I do use a skill that is more generic. Particularly when something is too new to have made it into the LLM training data set. Or not common enough.

I don't understand how "agent-browser" works.

Is it just the instructions? Where is the browsing executed? Locally with pupetter? Or it uses some service?

it's basically a cli for controlling a browser. the idea is that an agent like claude code would use it for validating something that it just did like changing something on the UI

What browser? My question comes from security, adding that skills just provides a line of bash, with no further info. I checked the .md file but it just lists a list of commands with agent-browser.

agent-browser is built on top of Playwright. Playwright uses a version of Chromium.

I was looking at this earlier. Has anyone used it? Is it useful compared to the Playwright MCP or Claude's Chrome plugin?

Agent browser is more lightweight than playwright mcp. Claude Chrome requires some manual setup, and works better in cases requires your actual browser not a headless one.

For the next model training version, would it make sense to incorporate all of these in the base model?

Not all. In fact a small model that has none of them but loads them on demand might be the most efficient thing

Official according to who?

Santa Claus

the skill finder meta skill, if it works, seems like a fantastic way to get untrusted prompts blindly injected into your agent

Honestly anything calling itself the “official” solution to Skills at this point is a scam at best.