People used to "upgrade" AI tools manually. Now AI agents increasingly do it themselves.
Not long ago, developers shared prompt templates when they met. The conversation has changed: which skill should you install so the agent immediately follows the project's rules?
Behind that change is a larger shift. AI development tools are getting something that looks a lot like a package manager.
On January 17, 2026, Vercel founder and CEO Guillermo Rauch wrote on X that Vercel was releasing skills: "npm for AI skills."
npm is the package manager frontend developers use every day. One command, and someone else's reusable code becomes part of your project.
Rauch was pointing to a similar mechanism for AI agents. Instead of copying a prompt from a post or repeating instructions from memory, you install a reusable capability package.
Peter Steinberger quickly replied that it would be worth syncing with ClawHub.
ClawHub is a skill marketplace from another agent ecosystem, unrelated to Vercel. The reaction was still telling: different AI tools are starting to converge on shared infrastructure.
A few days later, Vercel officially announced skills, a CLI tool for installing and managing packages of agent capabilities.
By July 6, 2026, the official vercel-labs/skills repository had reached about 25,000 GitHub stars.
Why did it grow so quickly?
It comes down to one command:
npx skills add <skill>
In practice, this is a package manager for AI agents. Like npm for frontend developers, but the installed artifact is not only code. It can be instructions, standards, references, templates, scripts, or a way of working.
And it is not tied to one tool. Claude Code, Cursor, Codex, GitHub Copilot, Gemini CLI, Windsurf, and many other agents can participate in this pattern.
Vercel also launched skills.sh, a catalog with install rankings. You can see which skills are popular, who published them, and how many times they were installed.
The leading skill is find-skills, with millions of installs at the time of review.
AI tools now have something close to a download chart.
One command, and the agent learns a new craft
In practice, the workflow is almost boring.
You type:
npx skills add vercel-labs/agent-skills
Press Enter.
A few seconds later, Claude Code has engineering standards for React and Next.js plus design guidelines. The next time it writes code, it can follow those rules.
Officially, this is a skill.
Technically, it is a folder built around SKILL.md: a file with metadata and instructions. It can also include scripts, templates, assets, or references. The agent reads the skill only when it is relevant, which keeps context cleaner than pasting the same long prompt into every session.
Why prompts were not enough
Prompt libraries solved an early problem: how do I phrase a task so the model behaves better?
Skills solve a different problem: how do I package repeatable behavior so an agent can discover and apply it inside a real workflow?
The difference matters. A prompt is usually copied by a human. A skill can be selected by the agent when the task matches its description.
This is a more scalable model:
- the project can define coding rules once;
- a team can share workflows;
- domain knowledge can live near the repository;
- scripts and references can be bundled with instructions;
- agents can apply the right behavior without the user pasting everything again.
Skills turn AI agents into configurable workers
An AI coding agent without skills is flexible but forgetful. It can solve many tasks, but every project starts by explaining the rules again.
With skills, the agent becomes configurable:
- this is how we write React components;
- this is how we review SQL;
- this is how we publish blog articles;
- this is how we handle design QA;
- this is how we deploy this service.
The important point is not that the agent becomes smarter in a general sense. It becomes better adapted to a local environment.
Why developers liked the idea
Developers already understand package managers. A package manager is a social and technical primitive:
- it gives a unit of reuse;
- it gives a distribution channel;
- it gives versioning and discovery;
- it gives a habit: install, inspect, update, remove.
Skills reuse that mental model for agent behavior.
Instead of asking "what prompt should I paste?", the developer asks "what capability should this agent have for this task?"
The safety problem
Skills are powerful, so they also need caution. A skill can influence how an agent reads files, edits code, runs scripts, and makes decisions. Installing random skills blindly is risky.
Teams should review:
- what instructions the skill gives;
- whether it includes scripts;
- what files it expects to touch;
- whether it changes security assumptions;
- who maintains it;
- whether it matches the project's style.
For production-adjacent repositories, skills should be treated like dependencies. Useful, but not magic.
What this means for Claude Code and Codex
Claude Code and Codex-style tools are no longer just chat windows pointed at a folder. They are becoming configurable execution environments.
The next layer of competition may not be only "which model writes better code." It may be:
- which tool has better skill discovery;
- which ecosystem has reusable workflows;
- which teams can encode their operating knowledge fastest;
- which agents can apply local rules without bloating every prompt.
That is why a skill marketplace matters. It is infrastructure for agent work.
The practical takeaway
If you use AI coding tools every day, start small:
- Identify one repeated instruction you give the agent.
- Turn it into a local skill or reusable workflow.
- Test it on a small task.
- Review the output.
- Share it with the team only after it proves useful.
The future of AI agents is not only larger models. It is also better memory for how a team actually works. Skills are one of the first practical forms of that memory.