Skills are the new app store.

The first wave of LLM apps looked like SaaS. Sidebar. Three-tab nav. A pricing page with three plans, the middle one in a gradient box.

The second wave is going to look like an app store inside a model. Not a webpage. Not a chat. A registry.

You'll discover a new "tool" the way you used to discover an iOS app — except it'll get installed into the agent you already use, not onto your phone. You won't open it. It'll just be available the next time the agent reaches for a verb it didn't have yesterday.

What a skill is, in 2026

A skill, in the form that's converging fastest, is roughly:

• A name, a description, a few examples
• A schema of inputs and outputs
• A short prompt teaching the agent when to invoke it
• An implementation behind a stable URL

The exact shape varies (OpenAI's Plugins, Anthropic's Skills, MCP, every framework's "tools") — but the underlying thing is the same: a verb you can teach a model to use, distributed independently of the model itself.

The skills market is going to look nothing like the apps market. There won't be a million skills. There will be tens of thousands of useful ones, and almost all of them will be boring — the equivalent of curl and grep rather than Photoshop. A skill for "find me the active Linear ticket for this Slack thread." A skill for "draft a release note from this PR diff." A skill for "summarize last week's Calendly bookings into a Notion page."

These won't have marketing pages. They'll have one-line descriptions and a ▶ install button inside whatever agent surface you happen to live in.

Why this is the market that gets eaten

A normal productivity SaaS today is competing for one of two things:

1. Attention — getting you to open a tab.
2. Workflow position — getting you to send your data through it.

Skills compete for neither. They compete for being the verb the agent reaches for. The user never opens them. The user never sees the brand. The user describes an intent and the skill is invoked, often without name, often without recognition.

This sounds like commoditization, and in a sense it is. But it's also the largest possible distribution channel that has ever existed, because the install action is free, instant, and triggered by the model itself the first time it needs you.

The companies that win this layer will look much less like SaaS and much more like infrastructure. Stripe, not Notion. Twilio, not Asana. Quiet, deeply integrated, paid per call.

How to position for it

Three things I've started doing:

1. Publish skills, not screens. Every new internal tool I build, I ship the agent-facing version first, and only build a UI if a human actually needs to look at it. About 60% of the time, no UI is required.

2. Treat the skill description as the brand surface. It's the only piece of marketing the agent ever sees. Write it like a tweet, not like an API doc.

3. Bias toward stateless and idempotent. Skills get invoked in unpredictable order by something that doesn't understand your domain. Anything that requires "but you have to do X first" will lose to a skill that doesn't.

The next decade of productivity software is going to look strange to anyone whose mental model still assumes a tab. The interface for getting work done is collapsing into the model. The thing you ship to participate in that interface is a verb.

Build verbs.