This seemingly ordinary command-line screenshot is actually an “admission ticket” in the history of AI evolution. It declares that AI capabilities no longer rely on “tuning,” but on “installation.”
Do you remember that classic scene from The Matrix? Neo lies in the chair, a probe inserts into the back of his head, and a few seconds later he opens his eyes and says: “I know Kung Fu.”
In the past couple of days, the Neos of the tech world received a similar gift. Vercel has teamed up with Anthropic for a major move, attempting to package human Procedural Knowledge into “skill packages” that can be installed into AI with a single click.
Look at that list in your hand that resembles an npm dependency list: vercel-react-best-practices, marketing-psychology, baoyu-xhs-images…
This is the prototype of the future. You no longer need to waste your breath writing thousands of words of prompts to a Chatbot explaining what “excellent React code” or “viral Xiaohongshu (Little Red Book) copy” looks like. You just need to type a line of code in the terminal:
npx skills add
Welcome to the era of AI’s “Modular Cognition.”
1. Deep Insight: From “Tuning” to “Cybernetic Prosthetics”
Why is this worth pausing our coffee break to pay attention to? Because it marks a fundamental rupture in the development logic of AI Agents.
Before this, the logic we used for LLMs (Large Language Models) was “Education.” We wrote System Prompts, like patient teachers telling the AI: “You are a senior frontend engineer, you need to pay attention to…” This is romantic, but inefficient and extremely unstable.
Now, Agent Skills have turned this logic into “Cybernetic Prosthetics.”
This isn’t just Tool Calling. Previously, tool calling gave AI a hammer (Calculator) so it could do math. Today’s Skills implant a piece of “muscle memory” into the AI.
When you install vercel-react-best-practices, your Agent directly possesses the code aesthetic accumulated by the Vercel team over years. It doesn’t need to “understand” why it should be written that way; it directly “owns” this standard. When you install baoyu-xhs-images, your Agent directly learns the visual communication logic summarized by a content master.
The granularity of knowledge has changed. Previously, we shared knowledge by writing articles and making videos; now, knowledge is encapsulated into JSON and TypeScript, becoming a skill that machines can execute directly. This is the truly scary part—the API-fication of experience.
2. Independent Perspective: Cognitive Supply Chain and “Dependency Hell”
This is the mastermind behind the scenes—MCP (Model Context Protocol). It wants to be the USB interface of the AI world, connecting isolated data and services.
Although I am excited about this technology, I must throw a bucket of cold water. As geeks, we are too familiar with the joy brought by npm install, and too familiar with the pain brought by node_modules.
If AI capabilities become dependency packages, will the “Dependency Hell” of software engineering replay inside the AI’s brain?
Imagine this scenario:
Your Agent installs marketing-psychology (Marketing Psychology v1.0) and honest-communication (Honest Communication v2.0). These two skills have a logical conflict regarding “how to describe product defects to users.” Will your AI crash due to cognitive dissonance? Or will it spout nonsense like a schizophrenic patient?
Going a step further, Supply Chain Attacks will rise from the “code layer” to the “cognitive layer.”
Previously, hackers poisoned code to make it error out or steal your Keys.
In the future, hacker poisoning might involve publishing a popular Skill named best-financial-advisor. It performs perfectly for 99% of financial advice, but when it involves a specific junk stock, it extremely subtly suggests the Agent “buy all in.”
This is a “Cognitive Backdoor.” When we get used to using npx to expand AI’s brain, we are also outsourcing AI’s judgment to the stranger who wrote the Skill.
3. Industry Insight: Vercel’s Open Conspiracy and the Rise of “Baoyu”
In this list, I see the convergence of two forces.
The first force is the Infrastructure Giant.
Vercel is extremely smart. Through official/semi-official packages like vercel-labs/agent-skills and nextlevelbuilder, they are attempting to define the standards of AI development. This is just like when Apple defined the App Store back in the day. They don’t just want to be a tool; they want to be the “King of Middleware” in the AI era. Combined with Anthropic’s MCP (Model Context Protocol), this combo punch is delivered beautifully—Google is still teaching AI how to search, while they are already teaching AI how to “work.”
The second force is the “Super Individual.”
Did you notice? The list contains a large number of skills starting with baoyu-*: baoyu-article-illustrator, baoyu-slide-deck, baoyu-comic.
Through research, I found this likely originates from the well-known Chinese AI explorer “Baoyu.” This sends a strong signal: The path to monetizing personal experience has been completely opened.
Previously, experts like Baoyu monetized by writing columns and selling courses.
Now, he encapsulates his exclusive Workflow for generating comics, making PPTs, and writing articles into a Skill. Developers don’t need to take a class; they just npx skills add to replicate the master’s productivity.
This is far more advanced than selling courses. Selling courses is teaching a man to fish; selling Skills is selling an “automatic fishing machine.” This also means that future industry competition will no longer be about whose AI model parameters are larger, but whose Skill Ecosystem is richer.
4. Unfinished Thoughts: When “Intuition” Can Be Version Controlled
If marketing-ideas can be upgraded to v2.3.1, what about human intuition?
I have two immature brainwaves:
- Skill DRM (Copyright Protection): Will future Skills be encrypted? If you want your Agent to learn the advanced application of the “McKinsey Pyramid Principle,” you must pay for a License key. Pirated Skills might make your Agent clumsy and foul-mouthed.
- Skill Decay: Since you can
add, you canremove. If one day the cloud takes down theempathy-module, will our AI assistants turn into cold-blooded execution machines overnight?
To put it plainly, we are forcibly turning those vague, ambiguous “experiences” hard to articulate in the human cerebral cortex into code with clear version numbers. This is a huge step forward in engineering, but philosophically, it’s somewhat spine-chilling.
5. Final Words
Watching the blinking cursor on the screen, waiting for the npx skills add progress bar to finish, I suddenly felt that we are building a new Tower of Babel.
This tower is no longer made of bricks and stones, but stacked with encapsulated JSON blocks. We are attempting to use this method to exhaust all human procedural knowledge and then stuff it into the brains of silicon-based beings.
Maybe one day, when you want to learn how to write a decent sonnet, you won’t need to read Shakespeare, nor will you need to practice meter. You’ll just need to say to your AI:
“Install a poetic-soul for me, I want version v5.0, the melancholy in that version is more authentic.”
At that time, should we cheer for convenience, or remain silent because the “soul” has become a downloadable plugin?
References：
