The agent skill ecosystem is exploding. Vercel’s skills.sh registry, launched just three months after Anthropic introduced Agent Skills in October 2025, hit one million skills in only seven months. This growth rate crushes traditional developer platforms: GitHub took 27 months to reach one million repositories, while npm required over nine years to hit one million packages. With nearly 280 million installs recorded, the registry has become the primary distribution channel for reusable AI instructions.

Supply vs. Demand: The Technical-Operational Gap

A deep dive into the registry’s aggregate data reveals a stark disconnect between what developers build and what businesses actually use. Supply is heavily technical, with software engineering alone accounting for about a quarter of all listings. However, demand is far more distributed. While software engineering remains the largest category for installs at 18%, business operations and writing each capture nearly 11% of the market. The data shows that business operations skills receive 74% more installs per listing than the average, indicating that non-technical teams are aggressively adopting agents for core workflows.

Workflow Automation and Concentration Risks

Agent workflow and automation skills are the second most installed category, representing 14.8% of the dataset. These skills teach agents how to plan, route work, and use tools, effectively allowing users to improve the agents themselves. Install activity is highly concentrated, with the top 1.2% of skills capturing 94% of all installs. Despite this concentration, there is no single winner; even the most popular skill accounts for less than 1% of total installs. Cross-industry skills, which apply to tasks like spreadsheet cleaning or website deployment, dominate with 87.5% of installs, proving that portability beats niche specialization.

The Shift to Proprietary Judgment

Vercel’s report suggests the first million skills taught agents general knowledge, but the next wave will focus on proprietary, company-specific judgment. As public skills and better models make general expertise a baseline, competitive advantage will shift to internal processes, such as specific refund policies or shipping criteria. The metric for success is also changing; while install counts currently signal quality, the future will demand rigorous testing and benchmarks to prove that a skill actually improves agent performance compared to using a raw model.

Key Takeaways

  • skills.sh reached one million skills in 7 months, significantly faster than GitHub or npm.
  • Software engineering dominates supply, but business operations and writing drive high install-per-listing ratios.
  • Cross-industry skills account for 87.5% of installs, highlighting the value of portability.
  • The top 1.2% of skills capture 94% of installs, though no single skill holds more than 1%.
  • Future value lies in company-specific judgment and effectiveness benchmarks, not just popularity.

The Bottom Line

The era of generic AI assistants is ending; the real competitive moat is now proprietary workflow data encoded into skills. Developers must pivot from building general utilities to creating deeply customized, testable skills that capture unique organizational judgment.