The 700,000-Skill Gold Rush: How 'Agent Skills' Quietly Became the New App Store

Ask three developers what an "agent skill" is and a year ago you'd have gotten four answers. It was a prompt. No, it was a config file. No, it was a folder with a markdown file and some scripts. The confusion was real, and it was the kind of confusion that usually kills a category before it starts.

Instead, the opposite happened. As of this month, the public directory at skills.sh lists north of 700,000 skills - reusable bundles of procedural knowledge that you install into an AI agent the way you'd install a VS Code extension. The leaderboard reads like the App Store top charts circa 2010: a couple of breakout hits with millions of installs, a long tail of niche utilities, and a surprising number of weekend projects that quietly went viral.

So what is a skill, really?

Strip away the marketing and a skill is almost boringly simple: a folder with a SKILL.md file describing what it does and when to use it, optionally bundled with a few helper scripts or reference docs. The agent reads the short description by default and pulls in the full instructions only when the task calls for it. That progressive-disclosure trick is the whole game - it keeps the agent's context window lean while giving it deep, on-demand expertise.

The format is deliberately dumb, and that's why it spread. A skill that teaches an agent how your team writes React components is the same shape as one that teaches it to redline a contract or generate a Flux image. Write it once, publish it, and anyone running Claude Code, Cursor, Copilot, Windsurf, Gemini or a dozen other agents can install it.

The land grab nobody announced

Here's the part that matters for the people building these tools. Whoever owns the install owns the relationship. That's why every major coding agent now ships first-class support for skills, and why the directory tracks installs per agent like a sports league tracks standings. The skill is portable; the loyalty is not.

• Claude Code leaned in hardest, treating skills as a native primitive rather than a bolt-on.
• Cursor's old "rules" files turned out to be skills in everything but name, so the migration was mostly cosmetic.
• Copilot, Windsurf, Cline and the rest followed, because the alternative was watching their users assemble their workflows somewhere else.

The breakout hits tell the story. A meta-skill called find-skills - a skill whose only job is to help an agent discover other skills - has well over a hundred thousand weekly installs. Anthropic's own frontend-design skill became a default starting point for anyone building a UI. Whole "packs" of media-generation skills riding on top of services like RunComfy show up again and again near the top of the charts.

Why this is good news for everyone (including us)

The genuinely useful thing about a 700k-skill ecosystem is that the boring, repetitive parts of working with an agent are now a download instead of a copy-paste ritual. You stop pasting the same "here's how we write tests" instructions into every session and start treating that knowledge as a versioned, shareable artifact.

That shift is exactly why we built our own Skills Directory. It started as our internal collection of IDE rules - the kind of "act as a senior full-stack developer, here are our conventions" prompts teams quietly hoard - and we've turned it into a proper, searchable catalog you can browse by stack and by agent. Think of it as a curated front porch to the same ecosystem: fewer entries, more signal, and everything one click from copy-and-paste.

The risk hiding in the long tail

A directory this big is also an attack surface. A skill is instructions plus, sometimes, scripts - and an agent that blindly trusts a skill is an agent you've handed to a stranger. The healthy ecosystems will be the ones that treat skill installs with the same suspicion mature package managers eventually learned to apply to dependencies: provenance, review, and a default assumption that popular does not mean safe.

For now, though, the gold rush is on. The models get most of the headlines, but the quiet truth of 2026 is that the leverage has moved up a layer - to the skills you teach them.