where to find ai skills

Someone posts a skill on a Friday. You read the description, think that is exactly the thing I keep doing by hand, and save the link. On Monday you open your terminal, find the link, copy the folder into place, and it works. Good week.

Three months later you're in a different tool. The skill isn't there. You half-remember who made it. You search, find something with the same name, and can't tell whether it's the one you used or a fork of it from before the author fixed the bug you never knew about.

That gap, between finding a skill and having it, is the part the search results don't cover.

what a skill actually is

A skill is a folder with a markdown file in it. The file starts with a name and a description of when it should be used, and then contains the instructions: how to do one job properly, written for a model rather than a person. It can carry supporting files (scripts, templates, checklists, reference tables) that the model reads only when it needs them.

That's the whole format. It isn't a plugin, it isn't code, and it isn't a prompt you paste. It's a document your AI reads and then follows. The reason it matters is the reason documentation matters: the judgement lives outside the conversation, so you stop re-explaining it.

Most tools that support skills read them from a folder. Some read several folders, including each other's. The format is open, which is why the same file can work in more than one place. If you want the longer version, what is a SKILL.md goes through it field by field.

the four places people look

Official repositories. The companies building these tools publish their own skills as worked examples: document handling, brand styling, testing, design review. They're well made and they're the closest thing to a reference implementation of the format. They're also narrow by definition: a company publishes the skills that show off what its own product does. Anthropic's skill creator is a good example of the genre.

Company repos. Infrastructure companies now ship skills the way they ship SDKs. If you work with a particular database, framework or deployment platform, the people who built it have probably written the skill that teaches an agent to use it correctly. These are usually the highest-quality skills available for anything technical, because the author has an interest in your agent not getting it wrong. Supabase's postgres best practices and Vercel's react best practices are two.

Community collections. Individual makers publish sets: a debugging methodology, a planning discipline, a writing review. This is where the genuinely interesting work is, because it's someone encoding how they do something rather than how a product works. It's also the least discoverable, because it lives wherever that person happened to put it. Jesse Vincent's systematic debugging is the kind of thing that only exists because one person wrote down how they work.

Directories and indexes. Sites that crawl public repositories and list what they find. Some list a few hundred; some claim millions. They're useful for exactly one thing: realising a skill for your problem already exists. They're less useful for the next step, because an index of everything is not a judgement about anything.

All four are worth using. If you're looking for a skill right now, start with the company that made the tool you're using, then the community collections, then a directory to catch what you missed.

the part that breaks

Here's what all four have in common: they end at the download.

Every one of them hands you a file and leaves you with a copy. That copy sits in a folder inside one tool. And from that moment three things start going wrong, slowly.

It only exists where you put it. Open a different assistant, a different editor, a different terminal, and the skill isn't there. You install it again, or you don't bother. Most people don't bother, which is why most people's skill folders have four things in them and the rest of their work runs on prompts they retype.

It stops being the version the author maintains. The maker fixes something in October. Your copy is from August. Nothing tells you. The skill still runs. It just runs the old judgement, and it will keep doing that for as long as you have it.

You can't tell whether to trust it. A skill is instructions your agent will follow, with your files and your access. The question that matters before you install one isn't does this look useful. It's who wrote this, and when did they last touch it. An index of two million entries cannot answer that, and mostly doesn't try.

None of this is anyone's fault. Installing into a folder is the simplest thing that works, and it was the right call when people had one tool and two skills. It stops being right at the point where you have four tools and twenty skills, which is roughly where a lot of people now are.

what finding a skill should mean

The useful version is closer to how you already handle a document you share with someone: you send a link, not a file. They open the link and see the current version, because there's only one. When you change it, what they see changes.

Applied to skills, that means a few things.

A skill has an address. One link, not a folder of copies, so the version you use is the version its author keeps.

The link works where you work. The same skill loads in the assistant you write in and the editor you build in, because the skill lives outside both of them rather than inside one. More on that in using one skill in every ai tool.

It carries its provenance. Whose it is, what it's for, when it was last tended. Enough to decide before you hand it your files.

It sits with everything else your AI knows about you. A skill is a way of working, which makes it the same kind of thing as your project context, your decisions and your preferences: the things you'd otherwise re-explain every session. Keeping them in separate places is a habit from when they had to be.

where to start

If you want to see what this looks like rather than read about it, the skills kept on innernet are open to browse. Some are made by us, some by the companies that build the tools, some by people who published theirs and kept them. Each one is a link. Paste it into any AI you've connected, and it loads there, at whatever version its maker keeps now.

browse the skills →

questions

where is the best place to find ai skills?

Start with the company that makes the tool you use, since its official skills are the reference for the format. Then look at skills published by the companies behind your database, framework or host, then community collections, then a directory to catch what you missed. innernet's community page gathers all three kinds in one place, each with its author, its source and the date it was last changed.

are ai skills free?

Almost all published skills are free and openly licensed, usually MIT, Apache or Creative Commons. A few official ones are source-available only, which means you may read and use them but not copy them elsewhere. Check the licence in the skill's folder or manifest before you redistribute one.

is it safe to install a skill from the internet?

A skill is instructions your AI will follow with your files and your access, so treat it like code you are about to run. Read it first, check who wrote it and when they last changed it, and prefer skills whose source you can see. A skill that tells the model to fetch and run something it doesn't show you is a reason to stop.

why does a skill I installed not show up in my other ai tool?

Because an installed skill is a copy in one tool's folder, and every tool reads its own. You can copy it into each tool, or keep the skill in one place every tool reads from, such as a link on innernet that any connected AI can load.

innernet is user-owned memory across AI tools, reachable over MCP. Connect it once and every tool reads the same thing: your projects, your decisions, and the skills you keep. innernet.live

Checked 17 september 2026. The skills landscape moves quickly; if something here has gone out of date, tell us.