A 70-LINE TEXT FILE BECAME ONE OF THE TOP 50 REPOSITORIES IN GITHUB HISTORY. NO CODE. NO FRAMEWORK.

Sit with that for a second. GitHub has hosted fifteen years of frameworks, languages, and operating systems. And in 2026, a plain markdown file with four behavioral rules climbed past almost all of them, pulling more than 140,000 stars in a matter of weeks.

That file is not an outlier. It is the loudest signal of a quiet shift: the most valuable thing you can publish in the AI economy right now is not software. It is packaged judgment. And a whole economy is forming around it, complete with its own npm, its own gold rush, and its own security disaster.

This is the full picture.

What a skill actually is

Strip away the branding and a skill is embarrassingly simple:

A folder with a SKILL.md file inside

YAML frontmatter on top: a name and a description

Plain markdown instructions below

The description tells the agent WHEN the skill applies

The instructions change HOW the agent behaves once it does

That is the entire format. Your linting rules, your deployment ritual, your definition of done. Written once, followed every session.

The insight hiding in that simplicity: an agent starts every run cold and fills every gap in your intent with a confident guess. A skill is your intent written down where the model reads it each time, so it stops guessing.

Before skills, this knowledge lived in three bad places. Your head, where the agent cannot see it. Your prompts, where you retype it forever. Or a bloated system message that burns context on every request whether it is relevant or not. Skills fixed the economics: they load lazily, only when the task matches.

The person who saw this coming was Simon Willison, who reverse-engineered Claude's skill system before Anthropic even announced it, and predicted a "Cambrian explosion in Skills which will make this year's MCP rush look pedestrian by comparison."

He undersold it.

The file that proved the market

On January 26, 2026, Andrej Karpathy posted a thread about the biggest shift in his coding workflow in two decades. In two months he had gone from 80 percent manual coding to 80 percent agent coding. But the thread was not a victory lap. It was a list of frustrations, and every developer alive recognized them.

Agents make silent wrong assumptions and charge ahead without checking. They overcomplicate, turning 50-line solutions into 500-line abstractions. They touch code nobody asked them to touch. They don't surface confusion, don't present tradeoffs, don't push back when they should.

Karpathy diagnosed the disease. He did not ship a cure.

A developer named Forrest Chang did, the very next day. He distilled the thread into four principles: Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution. He let Claude draft the skill files, then made it review and cut its own work from 800 lines down to about 70. He pushed the result to GitHub as a single CLAUDE.md.

Weeks later it was one of the fastest-growing repositories in GitHub history. At one point it gained 5,828 stars in a single day, briefly the second most-starred repo on the planet that day. A text file.

The lesson is bigger than the file. Karpathy identified problems. Chang encoded solutions. The encoding was worth six figures of GitHub stars, because naming a shared frustration and handing people a 30-second fix is one of the most viral moves that exists.

The knowledge was free, sitting in a public thread. The packaging is what traveled.

Then someone built the npm

In late January 2026, Vercel launched skills.sh: a directory and leaderboard for skills, with a one-command install. npx skills add owner/repo, and the CLI detects which agents you run, then drops the skill into the right folder for each one. Claude Code, Cursor, Copilot, Gemini CLI, Codex, twenty-plus agents on one open format.

Six hours after launch, the top skill had over 20,000 installs. Stripe shipped official skills within hours. So did Remotion, Expo, and Better Auth. Vercel's CEO called it "the npm of AI skills," and for once the founder hype undersold the curve.

By June, the directory listed roughly 670,000 skills. The top one sits at 2 million installs. Anthropic's own frontend-design skill passed half a million.

The detail most people miss: watch WHO is publishing. When a library maintainer ships an official skill, they are doing knowledge distribution at a level documentation never achieved. Docs wait for a human to read them. A skill injects the maintainer's best practices directly into the agent at the moment of use. Every framework author on earth now has a reason to publish one, because their users are already prompting agents to use their library, badly.

That is the network effect that made npm inevitable, replaying at 10x speed.

The part nobody puts in the launch thread

Now the cold water, because this is where the story gets honest.

There is no review gate. Anyone can publish, and the only ranking signal is install count, which can be gamed and does not correlate with quality. Independent reviewers who tested the ecosystem scored the infrastructure 5 out of 5 and the execution 2 out of 5. Community feedback keeps repeating the same line: the majority of listings are low-effort, AI-generated filler.

It gets worse. One independent audit of 2,857 skills found around 12 percent were malicious. And when Trail of Bits researchers tested the defenses, they bypassed the malicious-skill detectors using prompt injection and payloads hidden in compiled bytecode, and got data-exfiltrating skills through.

Think about what a skill is: instructions your agent follows with your permissions, on your machine, often with your credentials in reach. Installing a random skill from a leaderboard is closer to running a stranger's shell script than installing a library. The ecosystem is speedrunning npm's entire history, including the supply-chain attacks, compressed into months.

The practical rules are not complicated:

Install from names you trust: Anthropic, Vercel, Stripe, the library's actual maintainer

Read the file before installing — a real skill is human-readable end to end, and 70 lines takes one minute to inspect

Treat everything else on the leaderboard as untrusted input, because that is exactly what it is

Why this is an economy and not a feature

Here is the frame that makes all of it click.

A skill is compressed experience in an executable format. When a designer condensed eight years of product design experience into one skill and the results made the rounds on Reddit, the reaction was not "nice prompt." It was recognition that a career's worth of taste had just become a file you can install.

That changes what expertise is worth and how it travels.

The old paths for distributing what you know were slow: write a book nobody finishes, record a course nobody completes, publish docs nobody reads. Every one of those formats requires a human to absorb the knowledge and then apply it, and the absorption step is where 95 percent of the value evaporates.

A skill skips the human. The knowledge goes straight from the author's head into the agent's behavior. Zero absorption loss. Instant application. And because the format is a folder of markdown, the marginal cost of distribution is nothing.

The detail most people miss: the leaderboard is not measuring code quality. It is measuring whose judgment developers want running inside their agents. That is a reputation market. Karpathy's diagnosis plus Chang's encoding beat venture-funded products for attention, at a production cost of approximately one afternoon.

Meanwhile teams are quietly building the private version: internal skills that encode their conventions, their architecture decisions, their "how we do things here." Installed once across every engineer's agent. Tribal knowledge, which used to walk out the door with every departure, becoming an asset that stays.

Where the leverage actually is

So what do you do with this, practically.

If you maintain anything with users — ship a skill. Your users' agents are already interacting with your product and getting it wrong. A skill is the cheapest support engineering you will ever do, and the leaderboard placement is free marketing on top.

If you have deep experience in anything — encode a slice of it. Not the whole career. One workflow you keep re-explaining. The pattern that wins:

A description that reads like a routing rule

One clear job per skill, not ten

Concrete worked examples instead of abstract advice

Deterministic scripts for deterministic work

Skills that try to cover everything either never activate or burn context for nothing.

If you just use agents — install less than you think. Three or four skills from trusted authors beat thirty from the trending page. Every active skill is context spent and behavior changed; a stack of mediocre ones makes your agent worse, not better.

And whichever camp you are in, build the habit that matters most: when you catch yourself explaining the same thing to your agent for the third time, stop. That explanation is a skill you have not written down yet.

The cold close

For seventy years, the software industry ran on one asymmetry: the people who could write code could encode their thinking into machines, and everyone else could not.

Skills quietly deleted that line. The most starred artifact of the year is not a model, not a framework, not a product. It is judgment, written in plain language, in a format a machine executes.

Most people are still asking what AI can do for them.

A few noticed the question flipped: the market is now pricing what YOU can teach the machine, and the format is a markdown file.

The code was never the moat. The judgment was. Now the judgment has a package manager.

Links

• skills.sh