The first hundred skills on the shelf

A directory with a hundred skills on it is not a market. It is a promise with early evidence. The market — the part where installs compound, where an author's curve does the arguing so the author doesn't have to — comes later, if it comes at all. The first hundred publishers put their names on files before any of that existed. This is a note from the counter about what they showed up with, what they risked by going first, and what we owe them for it.

What shows up first

We did not seed the shelf with a taxonomy and wait for it to fill. We watched what people reached for, and the earliest skills clustered in three places.

Scaffolds came first — the project setup nobody wants to type twice. The Next.js foundation, the auth wiring, the schema and the seams between them. These are the workflows you have done enough times to hold a settled opinion about, and boring enough that you would rather hand them to an agent than narrate them out loud again.

Deploy rituals came next. The fragile sequence between "it works on my machine" and "it is live" — the k3s manifest, the environment check, the migration that has to run before the container flips. Nobody writes these down, because they live in muscle memory and a scrollback buffer. A skill is the first place that ritual stops being folk knowledge and becomes a file you can hand to someone else.

Then review taste, which is the hardest of the three. Prefer deleting a class to adding one. Extract a variant after the third repeat, not the first. These are the corrections you make on every pull request, the ones you stopped explaining out loud because explaining them takes longer than fixing them. They are the least likely to be written by anyone else, which is exactly why they are the ones worth signing.

The pattern under the pattern: the first skills are the workflows you run often enough to have taste about, and painful enough to want gone. Frequency plus friction. Where those two overlap is where the shelf fills first.

The description is the whole storefront

Here is a thing about this shelf that a human browsing it can easily miss: the buyer is not a human. When a skill gets installed, an agent fetches its manifest, reads the one-line description, and decides from that whether to pull the files. The description is not marketing copy hung above the product. It is the interface the installer actually reads.

{
  "type": "agent-skill",
  "name": "tailwind-refactor",
  "description": "Flattens divs, dedupes utilities, extracts variants.",
  "stats": { "installs": 3, "likes": 1 }
}

That reframes every naming argument the early publishers had with themselves. A clever brand name gives the agent nothing to match against a user's request. A verb gives it everything. Refactor, audit, rewrite — a model can line those up against "clean up my CSS" and know it has the right box in hand. The early publishers who rewrote their description more often than they touched the skill were not fussing over wording. They were tuning the one field the machine reads before it commits.

What going first costs

An empty shelf has no crowd to stand in. When you publish the hundredth skill there are ninety-nine curves around you; when you publish the third, there is you. No install count to vouch for the work. No retention number to point at. Nobody else's flat curve to look steady beside. You are the first data point, and the first data point carries all the risk of being wrong in public.

We are not going to wave that off. Putting your name on a file is a checkable claim, and early on the check comes back thin — a handful of installs, maybe one like, a published-at date whose only argument is that you were here before it was obvious. The reputational exposure is highest exactly when the evidence backing you is weakest. That is not a bug we intend to fix by faking the evidence. It is the actual shape of going first, and it is why the early names deserve more than the late ones.

The first hundred skills are not a sample of a market. They are the people who filled a shelf on the strength of a promise.

What the counter has learned

We built the counter to be boring on purpose, and that turns out to matter most at the beginning, when every number is small enough to lie loudly if we let it.

An install only counts when an agent fetches the file. The counter ignores human page views entirely, keys each install to a hash of the caller, and caps how many any single source can mint in a day. So an early number is small, but it is honest. There is no vanity traffic inflating it. Three installs means three environments that actually pulled the file into a project — not three hundred people who hovered over a card and moved on.

Which is why we tell early publishers to read direction, not total. Ten installs that climb week over week say something a thousand flat ones never will. Retention says more still: a skill that is still in the environment a week after it landed has cleared the only bar that finally matters. And the first installs have already confirmed the thing the original shelf-watchers suspected — narrow scope retains. A skill that promises one checkable thing and keeps it gets kept. A skill that promises a whole category gets deleted the first time it guesses wrong about which part of the category you meant.

None of that needs a big sample to show up. It shows up in the first dozen curves, because the mechanism, not the volume, is doing the work.

An honest accounting

So here is the ledger entry, stated plainly: a directory with a hundred skills does not yet have liquidity. The network effect — where every new skill makes the shelf more valuable to the next author, and every new author makes it more valuable to the next installer — has not switched on. The curves are short. The market is thin. Search still returns gaps. We are not going to dress that up, because a directory that lies about its own counter is worth precisely nothing, and the counter is the whole product.

What we can promise the people who went first is the part that compounds. The file never gets renamed — inbound links and installs outlive every rewrite, so the address you publish at is the address forever, and the reputation accrues to a stable place. The receipts stay attached to your name, not ours. And the upside of an early, signed, narrow skill is asymmetric: it costs the same to publish as a late one, and it accrues installs for longer, from a lower starting line, while the shelf around it is still filling in.

The hundredth publisher joins a directory. The first hundred built the thing that made a hundredth publisher worth becoming. That is the debt, and there is only one way a counter can pay it — by counting straight, keeping every address stable, and letting the curve say plainly who was early.

If you have been waiting for the shelf to fill before you add to it, you have the risk backwards. The crowded spots are where the arguing is already over. The empty ones are where the compounding still lives. Publish the file. Put your name on it. Be early on purpose.