# Skills Are Now a Vendor Channel

> NVIDIA ships 437 agent skills through the same install path as a practitioner's. What a vendor skill is for, what it is not, and the questions to ask before an agent loads one.

Canonical: https://mikereams.com/writing/skills-are-now-a-vendor-channel

Published: November 7, 2026  
Author: Mike Reams (https://mikereams.com/about)  
Topics: [AI in the Architecture Practice](https://mikereams.com/writing/topics/ai-in-the-architecture-practice), [Architecture](https://mikereams.com/writing/topics/architecture), [Artificial Intelligence](https://mikereams.com/writing/topics/artificial-intelligence)  
Tags: AI, Agent Skills, Governance

A skill is a folder with a SKILL.md in it: instructions an agent reads when a task matches. The [AI Library](https://mikereams.com/ai) on this site installs mine with one command, *npx skills add solventarchitect/skills*. In November 2026 NVIDIA put 437 skills behind the same command — *npx skills add nvidia/skills* — mirrored every day from its product repositories, built to the same [Agent Skills standard](https://agentskills.io/), aimed at the same agents. A vendor with a product catalog now has a skills catalog. That is a distribution channel, whatever the README calls it.

## What a vendor skill is for

Install the product. Configure it the way its engineers meant. Run it without the three forum threads it used to take. A vendor skill is the documentation compressed into the shape an agent reads, and because it is mirrored from the product repository it ages with the product rather than with somebody's blog post. For that job it will beat anything a practitioner writes, including me. I have never read the cuOpt install guide and now I never will.

## What it is not

Judgment. A skill that knows how to deploy a model does not know whether you should, which of the four teams asking for one has the data to feed it, or what the application portfolio looks like after it lands. The practitioner skills on this site — capability mapping, dispositions, lifecycle checks — sit one layer up. They decide **what**; the vendor skill decides **how**. An agent carrying both is the useful case. An agent carrying only the second is a very fast way to install things nobody asked for.

## Four questions before an agent loads one

- **Who verified it?** "NVIDIA-verified" means NVIDIA checked NVIDIA's. That is worth something and it is not independence.
- **What can it run?** A skill is text, but the agent reading it has your shell and your credentials. Read it the way you would read a vendor runbook before handing it to an operator, because that is what it is.
- **Which version?** Daily mirrors move. Pin the skill like a dependency, and treat the catalog's own CLI minimum as part of the pin.
- **Where does it run?** The same catalog now advertises sandboxed agent execution. That is the vendor saying what [a dry run is not permission](https://mikereams.com/writing/a-dry-run-is-not-permission) said: a skill may propose, a boundary decides.

## The cheap tier, live

The same catalog offers free hosted model endpoints for development, which made it the natural place to put the [two-tier pattern](https://mikereams.com/writing/two-tiers-spend-the-expensive-model-only-on-judgment) where you can poke it. The [AI Library](https://mikereams.com/ai) now ends with a box: the site's own search finds the pages, a small hosted model answers from those pages only, and every answer cites them. No bigger model, no memory, nothing stored. It is [better retrieval, not a bigger model](https://mikereams.com/writing/better-retrieval-not-a-bigger-model), running on somebody else's free tier, which is the correct price for the routine tier.

## The standard is the point

The interesting fact is not that a large vendor published skills. It is that they published them to a format a practitioner can share and an agent can carry alongside anyone else's. One agent, both catalogs, and the contest is over which skill the agent reaches for on a given task, not which vendor won. That is the right contest. Just keep the practitioner's skill in charge of the question, and the vendor's in charge of the answer.