← Back to blog
AiAbout 6 min read

Four of Today's Top GitHub Repos Are Skill Libraries. Now We Have a Packaging Problem.

Published Nov 7, 2026
Four of Today's Top GitHub Repos Are Skill Libraries. Now We Have a Packaging Problem.

I pulled the daily GitHub trending page on the morning of November 8. Fifteen repositories. Four of them are collections of agent skills: anthropics/skills, google/skills, mattpocock/skills and cloudflare/security-audit-skill. A fifth, vercel-labs/agent-browser, is a browser automation CLI that ships a skill wrapper alongside the binary.

That ratio is the news. A format that barely existed a year ago now occupies more than a quarter of the trending list, and the repos are coming from Anthropic, Google, Cloudflare and one very online individual with about 60,000 newsletter subscribers.

Adoption is going fine. The interesting failure is that there are four mutually incompatible ways to install the same conceptual object.

What a skill actually is

A skill is a folder. Inside it, a `SKILL.md` file with YAML frontmatter holding a name and a description, plus whatever instructions, scripts and reference files the task needs. The agent reads the description, decides the skill is relevant, and loads the body.

That is the whole contract, and it is specified at agentskills.io. The format is deliberately boring. Anthropic's repository describes skills as folders of instructions, scripts and resources that Claude loads dynamically, and notes that the docx, pdf, pptx and xlsx skills in the repo are the same ones powering the product's document features, published under Apache 2.0 as a reference implementation.

Boring is the point. A format simple enough to write in ten minutes is a format that spreads.

Four install paths, one object

Here is where it gets messy. Take the four trending repos and look at how you are supposed to get the skills onto your machine.

Anthropic's repo is a Claude Code plugin marketplace. You add it with `/plugin marketplace add anthropics/skills`, pick `anthropic-agent-skills`, then install `document-skills` or `example-skills`. Or skip the menu and run `/plugin install document-skills@anthropic-agent-skills`. After that you trigger a skill by mentioning it: ask Claude Code to use the PDF skill to extract form fields from a file.

Google's repo uses a different mechanism entirely. `npx skills add google/skills` drops you into prompts for which skills, which target agents, and whether the install is project-level or global. Updating is `npx skills update`, which also offers to remove obsolete skills. The catalogue is large and organised by product: IAM recipes, GKE inference migration, BigQuery AI, Genkit in four languages, Live API, and a dozen separate skills for the Agent Platform alone.

Matt Pocock's repo gives you a choice and warns you about it. Use `claude plugin install mattpocock-skills@claude-plugins-official` and the plugin updates itself. Use skills.sh and you get editable files copied into your project that you update by hand. His README contains the single most useful sentence in this whole category: pick one per agent, because installing both gives you every skill twice.

Cloudflare's security audit skill goes direct. `npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit`, with `--global` for a user-level install. Then you point your agent at a codebase and ask it to do a security audit.

Four products, four mental models. A plugin marketplace with a browse-and-install UI. A registry-driven CLI with interactive prompts. A copy-into-your-repo model with manual updates. A single-skill pin by URL.

Why duplication is worse than it sounds

The obvious problem is disk space and clutter. The real problem is precedence.

When two skills with overlapping descriptions are both installed, the agent chooses based on the description text it reads at load time. It does not know that one is a stale copy of the other. It does not know that the project-level version was meant to override the global one. It reads two paragraphs that both sound relevant and picks.

That produces behaviour that is very hard to debug. The agent does the wrong thing competently. You go looking for a bug in the skill and find that a different skill ran.

The update story makes it worse. A plugin that updates itself can change behaviour under you between sessions. Copied files stay where they are until someone remembers to pull. And the `npx skills update` route can remove skills you were relying on, which is a good feature and a surprising one if you did not read the prompt carefully.

The trust problem nobody has priced

A skill is instructions an agent will follow, often with tool access, often with the ability to run bundled scripts. Installing one is closer to installing a package than to reading a blog post.

Cloudflare's repo shows what a serious one looks like. It runs a structured audit in six phases, from reconnaissance through coverage-led hunting, candidate validation, structured output, independent record verification, and a target-neutral report. It writes an architecture file and a coverage ledger, then validates them with two zero-dependency Node scripts, and it treats repeat runs against the same repo as additive, using prior ledgers to target gaps rather than re-covering old ground.

That is a lot of machinery, and it is the sort of thing you would want to read before handing it a codebase. Which is exactly the point. The install command takes four seconds. The review takes an afternoon.

There is no registry signature, no provenance metadata beyond the git remote, and no lockfile recording which version of which skill is active in which agent. For a format this popular, that is a gap.

The app layer is already adapting

One more item from the same trending page is worth noting because it is a different kind of signal. meta/astryx appeared at number six: an open source design system on React 19 and StyleX, currently in beta, described as fully customisable and agent ready.

Read that as a manufacturer noticing who the new customer is. If agents generate interfaces, the component library needs to constrain them. Tokens, variants, and rules that a model can follow are a different product from a system designed for humans to browse.

The same logic explains why skills are spreading. Someone has to encode the house style, the deploy checklist, the audit procedure, the way your team writes a migration. Once you write it down as a skill, the agent stops improvising.

What is missing

The category has a format, a spec, several registries, and four installers. What it does not have is the boring infrastructure that makes a package ecosystem survive: a lockfile, a resolved dependency graph across agents, signed provenance, and a clear rule for precedence when a project skill and a global skill collide.

Those will arrive, because they always do. In the meantime, the practical advice is the advice Matt Pocock already wrote into his README. Pick one installation path per agent. Know where your skills live. Read the ones that can run code.

The repos are trending because skills solve a real problem: agents are general and your work is specific, and a folder of instructions is the cheapest way to close that distance. The packaging chaos is the tax on finding out.

Related articles