Skills.sh Review: An Open Directory for AI Agent Skills

Your AI coding agent is smart, but it doesn't know your workflow. It doesn't remember the exact way you structure React components, the specific linting rules your team enforces, or that one custom deployment script that always needs special handling.

Enter Skills.sh.

Launched by Vercel in January 2026, Skills.sh is an open directory where developers share reusable "skills": procedural knowledge packaged as Markdown files that teach your AI agent how to handle specific tasks without cramming everything into the context window.

This review looks at whether it actually makes agents smarter or falls short of the expectations, focused on Claude Code and Cursor workflows.

What Skills.sh Actually Is (And What It's Not)

Skills.sh is a leaderboard and discovery platform for AI agent skills. Think of it as npm for agent capabilities, but instead of installing code libraries, you're installing procedural knowledge.

Below is the workflow:

Browse the leaderboard at skills.sh
Find a skill you want (like "vercel-react-best-practices" or "rails-7-conventions")
Run one command: npx skills add <owner/repo>
Your agent now has access to that skill.

What it is not:

Not a hosted service with subscriptions or rate limits
Not a standalone agent or IDE
Not a guarantee that community-submitted skills are high quality

It is open, free, GitHub-based infrastructure.

The One-Command Installation That Actually Works

This is where Skills.sh shines. With most agent enhancement tools, you're editing config files, copying prompts into system messages, or maintaining custom instruction documents that your agent may or may not read properly.

Skills.sh installation looks like this:

npx skills add vercel-labs/vercel-react-best-practices

The CLI detects which agent you're using (Claude Code, Cursor, GitHub Copilot, Aider, etc.), drops the skill into the right directory (such as .cursor/skills/ or .claude/skills/), and the next time your agent starts, it has access to that skill.

Typical usage looks like this:

Claude Code: Install React testing patterns, and the agent should start suggesting better test structure in PRs.
Cursor: Add Rails conventions, and Cursor should stop suggesting Express.js patterns in a Rails codebase.

The experience feels seamless, but its usefulness depends heavily on the skill quality.

The Agent Compatibility Promise (Multi-Platform Support)

One of Skills.sh's strongest features is its agent-agnostic design. The same skill works with:

Claude Code
Cursor
GitHub Copilot
Aider
OpenCode
Codex CLI
Any agent that reads from a standard skills directory

This matters because closed ecosystems lock you into a single vendor's internal prompts or context systems. With Skills.sh, you can switch agents and bring your skills with you.

How Skills Are Built (Markdown + Optional Scripts)

Skills are simple. Each skill consists of:

A Markdown file with instructions written in natural language
Optional shell scripts or templates
Optional assets (like example files)

Example structure:

my-skill/
├── SKILL.md          # The main instructions
├── scripts/
│   └── setup.sh      # Optional automation
└── templates/
    └── example.tsx   # Optional code samples

The SKILL.md file is what your agent reads. It might instruct:

"When creating React components, always use function syntax, define PropTypes at the top, and include error boundaries for any component that fetches data."

Your agent interprets this as procedural guidance, adapting the skill to the specific code it is writing. Natural language instructions instead of code are what make skills portable across different agents.

The Leaderboard (Discovery by Install Count)

Skills.sh ranks skills by install count and trending momentum. The site shows:

Hot: Skills gaining traction in the last 48 hours
Trending: Sustained growth over the past week
Top: Most installed overall

Telemetry is anonymous; only aggregate install counts are collected. The top skills are predominantly from verified vendors (Vercel, GitHub, Anthropic, Cloudflare), while community submissions are often buried under the noise.

The Quality Problem

Skills.sh currently lacks active quality control. Anyone can create a skill, host it on GitHub, and make it available for installation. The only ranking mechanism is install count, which does not always correlate with quality.

Feedback from developers highlights this challenge:

Developers note that a significant portion of trending community skills are overly generic, simply restating agent defaults.
Some trending skills contain contradictory advice that conflicts with official framework documentation.
Other skills reference deprecated APIs and are already outdated.

To get the most out of the platform, it is best to stick to vendor-provided skills and options from trusted community members. The platform would benefit from curation, verification badges, or community ratings.

The Privacy Question (Unknown Data Policy)

Skills.sh does not feature a published privacy policy or terms of service on its website. While the CLI collects anonymous telemetry (install counts), it does not detail:

How long that data is retained
Whether skill usage patterns are used to train models
What happens to your skill definitions if you delete them

If you work in a regulated environment (such as healthcare, finance, or government), you may need to evaluate these security aspects before introducing the tool to your team's workflow.

When Skills.sh Is a Great Idea

Skills.sh works best when:

You are using Claude Code or Cursor and want to extend their capabilities without switching agents.
You have team-specific conventions that agents frequently forget.
You want to avoid bloating your context window with repetitive instructions in every session.
You are comfortable vetting and testing skills yourself.

When Skills.sh Isn't the Right Fit

You might want to skip Skills.sh if:

You require enterprise-grade audit logs or single sign-on (SSO).
You want a fully curated and verified library.
You are looking for non-coding skills (the platform is strictly for developer workflows).
You prefer fully decentralized systems (it relies on central GitHub and Vercel infrastructure).

Alternatives to Skills.sh

If you are looking for alternatives in the agent skills space, consider these options:

Skills Directory: Focuses on verified, community-curated skills. It has a smaller library but a higher signal-to-noise ratio.
Agent Skills Marketplace: A massive repository with over 65,000 skills and daily automated updates, offering more advanced filtering options.
MCP Market Skills: Specializes in Claude, ChatGPT, and Codex enhancements with detailed categorization and installation guides.
Awesome OpenClaw Skills: A public GitHub collection for local AI assistants, providing a transparent, CLI-free alternative.
Cloudflare .well-known/skills Proposal: A conceptual standard for decentralized skill discovery via URIs.

Skills.sh vs. Built-In Agent Features

Feature	Skills.sh	Claude Code Built-In	Cursor Skills
Price	Free	Free	Free
Cross-Agent	Yes (any agent)	No (Claude only)	No (Cursor only)
Quality Control	None	Anthropic-verified	Cursor-verified
Privacy Policy	Unknown	Documented	Documented
Community Library	Yes	No	Limited
Enterprise SSO	No	Yes	Yes

Skills.sh trades quality control and corporate compliance for cross-agent flexibility and open accessibility.

The Verdict

Skills.sh is a powerful implementation of agent skill distribution. The one-command installation works smoothly, the cross-agent compatibility is highly functional, and the underlying architecture is clean.

However, the lack of quality control, missing privacy guidelines, and low signal-to-noise ratio in community-submitted skills mean developers must carefully vet what they install.

Final Score: 3.5 / 5