Vercel Finds Agent Skill Installs Concentrated in a Tiny Share of Its Registry

Its skills.sh report points to demand beyond coding, but registry install counters cannot show how many people chose a skill or whether it helped.

By 3 min read
Vercel Finds Agent Skill Installs Concentrated in a Tiny Share of Its Registry
Vercel Finds Agent Skill Installs Concentrated in a Tiny Share of Its Registry

Listen to this story

The audio brief

About 1:35
0:001:35
Read transcript
Just 375 skills in Vercel’s registry generated 62 percent of all installs. That’s a tiny slice of a catalog that passed one million listings and recorded nearly 280 million installs in seven months, according to the company’s new report on skills.sh. A skill is a set of reusable instructions that helps an AI agent handle a task. The report’s clearest signal is how unevenly people install them: nearly half of the listings were installed exactly once, while the top 1.2 percent captured 94 percent of installs. The pattern also challenges the idea that demand simply follows the biggest category. Software engineering makes up about a quarter of the catalog, but draws 18 percent of installs. Business operations and writing each draw nearly 11 percent, and their listings average more installs than the overall average. Skills that apply across industries show an even stronger gap: they make up 66 percent of classified supply, but 87.5 percent of installs, averaging 3.6 times the installs per listing of industry-specific skills. That may help explain why broadly useful instructions travel farther. But these are registry counters, not unique users, evidence of continued use, or proof that a skill improves an agent’s work. The category breakdown also comes from a classified sample, not the full catalog. Vercel expects tests and benchmarks to help measure whether an agent does a job better with a skill than without it. For now, the key constraint is simple: installs show what gets picked, not whether it works.

Story brief

3 key points

Vercel’s seven-month snapshot of skills.sh suggests that demand for agent skills favors broadly useful instructions, but it does not show whether those skills improve an agent’s work. Business-operations and writing listings attract more installs per listing than average, while software-engineering skills—about a quarter of supply—average roughly 30% fewer. The results may inform how builders package and promote...

  1. 01

    Just 375 listings—0.04% of the registry—generated 62% of installs; the top 1.2% accounted for 94%.

  2. 02

    Nearly half of listed skills were installed exactly once, underscoring the registry’s long tail.

  3. 03

    Cross-industry skills make up 66% of classified supply but 87.5% of installs, averaging 3.6 times the installs per listing of industry-specific skills.

A million skills give AI agents a vast menu of reusable instructions. Vercel’s new study of its skills.sh registry finds that installs are far less evenly spread: just 375 listings account for 62% of the total. The gap between what has been published and what gets installed is the report’s clearest finding.

Published September 25, the State of agent skills report says the registry passed one million listings and recorded nearly 280 million installs in seven months. A skill supplies an AI agent with instructions for a particular job. Vercel launched skills.sh in January as a directory where people can find and install them, after Anthropic introduced Agent Skills in October 2025.

These are reusable directions, not necessarily new models or standalone applications. Anthropic describes a skill as a folder that can hold instructions, scripts and other resources. Someone can use that format to record the steps and judgment behind a task for an agent to reuse. A listing can be as simple as instructions written in ordinary language, helping explain how the catalog can grow quickly.

Coding dominates the catalog, not the installs

Software engineering accounts for about a quarter of listings, Vercel says. Together with agent workflows, data, infrastructure and security, technical categories make up more than half the catalog. But in Vercel’s classified dataset, software engineering draws 18% of installs. Agent workflows account for about 15%, while business operations and writing each draw nearly 11%. No single category claims a fifth.

The difference sharpens when installs are measured per listing. Business operations skills average 74% more installs than the overall average; writing and document skills average 50% more. Software engineering skills average about 30% fewer. That comparison describes activity relative to each category’s supply, not the number of people using a skill or the results they achieved.

Why broad instructions travel farther

Skills for work shared across industries represent 66% of Vercel’s classified skills but 87.5% of their installs. They average 3.6 times as many installs per listing as skills tied to a specific industry. Vercel points to jobs such as cleaning spreadsheets or deploying websites: the same instructions can apply in many kinds of organizations.

Some skills travel across tasks as well as industries. Agent-workflow instructions teach an agent to plan, route work or use tools, rather than complete just one specialized job. Among skills with at least 100,000 installs, agent workflows are the largest category, Vercel says. A planning skill may have more chances to be useful than one written solely for contract review, though registry counts cannot establish which produces better work.

An install is not a result

The distinction limits how far to take the findings. Install figures come from aggregate registry counters; they do not represent unique people or necessarily independent choices. The function and industry breakdowns use a classified sample, not the full catalog. A large install count therefore shows reach within this registry, not how widely a skill is used after installation.

Vercel expects the measure of a skill to shift from popularity toward whether an agent does a job better with it than without it. The company predicts that tests and benchmarks will help answer that question, and that organizations will write more skills containing their own procedures and judgment. For now, the study maps what gets listed and installed; it does not measure the improvement those instructions deliver.

Sources

  1. vercel.comIntroducing skills, the open agent skills ecosystem - Vercel
  2. anthropic.comEquipping agents for the real world with Agent Skills
  3. vercel.comState of agent skills

Loading discussion...

YOUR READING SPACE

Notifications