The biggest change in AI this year isn't a new model. So what is it, and how can you take advantage of it?
Every time a major AI company has announced a product in the last six months, there's been a feature buried in the announcement that nobody in the tech press seems to have noticed. In October, Anthropic published an engineering blog post describing a new architecture for giving AI agents installable, reusable capabilities. The headline coverage was about Claude getting better at specialised tasks. But buried underneath was the actual news: Anthropic had just created a standard that would let any AI agent, on any platform, load permanent new capabilities.
By December 2025, Anthropic had open-sourced it. Within weeks, OpenAI had adopted it for Codex and ChatGPT, GitHub Copilot had added it to VS Code, and Vercel had built an entire distribution layer for it, which is now quietly doing a quarter of a million installs a week.
Anthropic called these shippable capabilities "agent skills". And it's the most important architectural shift in how AI gets used since OpenAI's ChatGPT release in 2022.
This post covers what a skill is, why the change matters more than it sounds, how fast the market around it has formed, and how to install your first one this afternoon.
What a Skill Actually Is
It's easiest to think about a skill as an app for your agent. In the same way you download an app for your phone on a marketplace (the app store), your agent can now install its own app (a skill) through a dedicated skill marketplace/registry (ClawHub).
Mechanically, a skill is a small bundle of files. It usually includes a text file called SKILL.md that tells the agent what the skill does, when to use it, and how it works. Some skills are just detailed instructions while others include code the agent can run, an API the agent can call, or a set of rules for handling specific kinds of input. A skill for reading invoices might include a Python script that parses PDFs. A skill for calendar management might include the logic for connecting to Google Calendar, checking your availability, and proposing meeting slots. A skill for summarising Slack threads might be detailed instructions on how to read a thread and produce a clean summary of your conversations.
Skills live in registries. The biggest public one is ClawHub, which currently hosts over 52,000 of them. When you install a skill from ClawHub, its files get added to a folder your agent reads from. The next time the agent is doing work and encounters a task the skill covers, it uses it.
Why the Architectural Change Matters
Before skills, an AI agent's capabilities were functionally fixed at whatever the model had been trained on, plus whatever you explained to it in the moment. If you wanted it to handle your invoicing, you had to write a careful prompt describing your invoice format, your accounting categories, where things should be filed, and how to handle edge cases. You had to do that every session, because the model forgot everything between conversations. If you wanted a different task, you wrote a different prompt. That was the limit on what anyone could get out of an AI agent without being an expert prompter.
Skills remove this limit. The capability is now a file someone has already written, tested, and published. You install review-invoices and the work has been done by whoever built the skill. Your agent handles invoices the way it handles every other installed capability: automatically, without instruction, in the background.
The smallest reusable unit of AI work used to be a prompt. Now it's a skill. Prompts are written once, by one person, and die at the end of the session. Skills are written once and reused by everyone who installs them. That's a different kind of scaling, and it's the same kind of scaling that made the App Store a different phenomenon from software you wrote for yourself.
The other shift is that skills work across platforms. Anthropic open-sourced the SKILL.md standard in December 2025, which means a skill written for Claude also runs on ChatGPT, GitHub Copilot, Cursor, and most other agent platforms you'd realistically use. That kind of cross-platform compatibility is rare, and it's what turns skills from a Claude feature into the beginning of an actual ecosystem.
How Fast This Is Actually Moving
The other thing worth noting is the pace.
The SKILL.md standard is four months old. In those four months, OpenAI adopted it for Codex and ChatGPT, GitHub Copilot added it to VS Code, and Cursor implemented it. Vercel launched skills.sh, a cross-platform skill installer, in January; one of its utilities, find-skills, is currently doing over 235,000 installs per week. ClawHub has grown to more than 52,000 community-published skills. Agensi, the first paid marketplace for skills, launched at the beginning of the year and already hosts over two hundred paid skills with live payments flowing to creators.
The Apple App Store took six months to reach 15,000 apps. npm, the JavaScript package manager half the internet runs on, took around eight years to cross 350,000 packages. The wider agent skills ecosystem crossed the same threshold in about two months.
This is not a slow-moving market forming gradually. This is a market forming in real time, faster than any comparable technology shift in recent history, and almost nobody outside the ecosystem has noticed yet.
What People Are Doing With Them, and Where the Market Is Going
The simplest use is individual workflow automation. An operator who used to spend an hour every Friday reconciling invoices can now install a skill like an Invoice Extractor, which parses invoices, extracts line items, and logs them to the right accounts. The hour of work becomes a minute. This, multiplied across the dozens of repeatable tasks in a typical freelance or small-business week is a huge compression of the working day.
The more interesting use of skills is how people are packaging that automation and selling it as a service. A one-person operation running a customised skill stack for invoice processing, lead enrichment, research-to-publish, or any of a dozen other repeatable functions can now deliver results that would previously have required a small team. That's a new kind of opportunity, one that didn't exist a year ago. We'll cover specific winning workflows you can use in upcoming Stack Plays.
Then there's where the market itself is going. Most skills today are free. That won't be permanent. Agensi launched a paid marketplace at the beginning of the year and is already processing payments to skill creators. Skills4Agents is in waitlist with an even more creator-friendly revenue split. The commercial infrastructure for paid skills is being built right now, and the window in which skills are free as the default is beginning to close.
Who Writes Skills, and Why Not All of Them Are Safe
Anyone can publish a skill on ClawHub. Most are useful, some are weekend experiments, and a small number are deliberately malicious. That's the trade-off of an open registry. Earlier this year ClawHub had a supply-chain incident nicknamed "ClawHavoc", in which over 1,400 malicious skills were found and removed. The registry has since added security review, but the open model means harmful skills will keep appearing. Spotting them before you install is a skill in itself, and one we'll keep covering.
How to Install Your First Skill
The full walkthrough, covering every supported platform, the exact commands, and what to do when something doesn't work, is in a companion post: How to Install Your First Agent Skill (Start to Finish).
When it works, you'll have installed a capability into an AI agent. Most people who use AI every day haven't done that yet, and this is about to be the default way people use AI for everything.
Agent skills are new. The market is young. The opportunity is clear. This is the only publication mapping the emerging skill economy and teaching you how to capitalise on it. Subscribe to stay early.
This post is part of The Skill Economy, a weekly look at the AI agent skills market. If you want to understand this space before the rest of the world catches up, subscribe for weekly analysis, teardowns, and workflows.
Top comments (0)