In August last year, about forty people a month typed “agent skills” into Google in the United States. Anthropic published the format that October. By January this year the phrase drew 18,100 searches a month, and it sits at 8,100 today. I pulled the full thirteen-month series on 24 October 2026 through Ubersuggest.
Then I looked at who owns the first page for it. GitHub, Microsoft Learn, the Claude platform docs, Elastic, Sysdig, and a spec site called agentskills.io. Engineers wrote every one of those pages, for engineers. Search the comparison version, “agent skills vs mcp”, and Google puts an AI Overview above everything. Below it sit LlamaIndex, Red Hat and a Reddit thread.
Not one of them speaks to somebody who runs a business alone and has never opened a terminal. That gap is worth closing, because the idea underneath the jargon involves no code at all. What it asks you to do is write down the thing you have been retyping into every chat.
The clearest example I have is one I run every week. I keep a list of the phrases I will not publish, and a draft gets a pass against that list before anything goes out, which usually means deleting about five AI tells. That list is the content skill from the examples further down, and it works because it is a file rather than something I try to remember at the end of a long day.
What are agent skills?
An agent skill is a folder containing a written instruction file that your AI reads by itself when a task matches it. The folder contains one file called SKILL.md, which says what the skill is for and how you want the work done. Templates and examples can sit alongside it. Your AI reads the short description of every skill you have and notices when one is relevant. It opens that skill and follows the instructions inside. Nobody has to remember to paste anything.
Anthropic introduced the format in October 2025. Its engineering team describes skills as “organized folders of instructions, scripts, and resources that agents can discover and load dynamically”. The Claude platform documentation puts the same idea in plainer words. Skills there are “modular capabilities that extend Claude’s functionality”, and “each Skill packages instructions, metadata, and optional resources (scripts, templates) that Claude uses automatically when relevant.”
Who controls the format now?
The format was then released as an open standard, so it is no longer one company’s feature. Anthropic’s own post has an update dated 18 December 2025 saying it had “published Agent Skills as an open standard for cross-platform portability”. The specification now lives at agentskills.io, which describes agent skills as “a lightweight, open format for extending AI agent capabilities with specialized knowledge and workflows.”
The spread since then is the number worth keeping. When I counted the entries on that site’s client showcase on 24 October 2026, forty-six agent products were listed as supporting the format. They include Claude and ChatGPT alongside Codex, Claude Code, Cursor, GitHub Copilot, VS Code and Gemini CLI, which means a skill you write once is not tied to one subscription.
Two other things share this name. Agent skills here are not the professional abilities of a real estate, insurance or talent agent. That is the meaning behind most of the older search results. They are also not the character upgrades called Agent Skills in Digimon Story: Time Stranger. That video game shares the name, turns up in the results for it, and has nothing to do with AI.
How does a skill actually work?
Your AI reads the one-line description of every skill you have. It opens the full file only when a task calls for it. That mechanism has a name, progressive disclosure. The Claude documentation defines it as a case where Claude “loads information in stages as needed, rather than consuming context upfront.”
The agentskills.io specification breaks it into three stages, which it calls discovery, activation and execution. At startup the agent loads only the name and description of each skill, which is enough to know when it might be relevant. When a task matches that description, it reads the full instructions. Then it follows them, opening any bundled templates or reference files it needs.
The practical effect is the part worth understanding. You can keep a dozen written procedures on hand without any of them cluttering the conversation. Your AI only opens the one that fits what you asked for. Anthropic’s engineering team makes the same point about scale. Agents with file access “don’t need to read the entirety of a skill into their context window”. The result, in their words, is that “the amount of context that can be bundled into a skill is effectively unbounded.”
Simon Willison reviewed the specification on 19 December 2025. He called it “a deliciously tiny specification” and said you can read the whole thing in a few minutes. He also called it “quite heavily under-specified”, which is fair and matters for the limits section below.
What is the difference between a skill, MCP, a prompt and custom instructions?
They operate at four different layers. People selling courses use the words interchangeably, so ninety seconds here saves you money later.
- A prompt is what you type this time. It applies to one message and disappears when the chat ends.
- Custom instructions are the standing rules that apply to every chat, everywhere. They are always loaded, so they have to stay short, and a procedure will not fit in them.
- An agent skill is a written procedure your AI opens only when it is relevant. It can be long, because it is not competing for space with everything else you have written down.
- MCP stands for Model Context Protocol. It connects your AI to live systems such as your calendar, your inbox or your database. Skills tell your AI how to do something, and MCP gives it something to reach.
The clearest distinction I have read comes from Clelia Astra Bertelli and Tuana Çelik at LlamaIndex, writing on 3 February 2026. Skills, they write, “act as contextual modification, by injecting custom instructions to the agent’s context”. MCP servers instead offer “a functional interface, which executes specific logic under specific conditions”. They add a line worth quoting to anyone unsure which to reach for. With an MCP tool the agent only decides which tool to run and when. With a skill, the challenge is “to decide which skill to use, when, and how.”
For a one-person business the answer is almost always skills first. You have far more undocumented procedure than you have systems to connect.
Do skills work if you do not write code?
Yes, and nothing about the format requires it. A skill is a folder with a markdown file in it. Markdown is plain text with a few hashes for headings. There is nothing to install, nothing to run and no repository to manage. The required fields amount to a name and a description at the top of the file.
Here is the part that tends to land. You almost certainly already have skills, written in the worst possible place. You retype “short paragraphs, no exclamation marks, end with a question” into every fresh chat. That is a skill, run from memory. You re-explain your newsletter structure, your offer, and the words you never use. Each time, you recite a procedure that should have been a file eleven months ago.
The reason your AI keeps producing average work is not that the model is weak. Everything it needs to know about how you work lives in your head instead. You type it out inconsistently, and it disappears the moment you close the tab.
What would skills look like in a one-person business?
Four examples, all buildable in an afternoon, none of them requiring code.
- A content skill. Your post structure, your hook rules, the phrases you refuse to publish, and two examples you liked. Your AI opens it whenever you ask for a caption.
- An offer skill. What each product is, who it is for, what it costs, and what it does not include. This is the one that stops your AI inventing features you do not sell.
- A client-reply skill. How you answer a pricing question, how you decline politely, and what you send after a first call.
- A repurposing skill. How a long piece becomes an email, and how that email becomes three short posts. Write it as steps rather than as a wish.
Notice what all four have in common. Each one is a thing you already know and have never written down, and the writing is the work. The folder takes two minutes.
A skill is also the thing a narrower helper needs if you ever hand one a job, because the helper gets instructions rather than your whole conversation.
That set of files has a proper name, and it is a knowledge base for AI. Skills are the procedural half of it, meaning how you do things. The other half is factual, meaning what you sell, who buys it and how you sound. Both halves are what Business Brain builds.
What will skills not fix?
Four honest limits, because a page that only lists upsides is not worth citing.
A skill will not decide what your business should say. It executes your thinking faster and more consistently, and the thinking stays yours.
A skill file inherits whatever quality you put in it. Vague instructions produce vague output. Writing “make it sound professional” into a file changes nothing.
The standard is young and still moving. Simon Willison’s “quite heavily under-specified” reading from December 2025 still applies, and a long client list is evidence of adoption rather than of stability. Support differs between tools, and something you write today may need adjusting later.
Skills go stale exactly like every other document. A skill describing last year’s offer will confidently produce last year’s content, and nothing in the format warns you.
Where do you start?
Start by finding out what your AI currently knows about your business. Most people are wrong in both directions about that.
I caught myself on this a few weeks ago. The way I use AI every day had started to feel ordinary to me, and then I watched my own team work and saw that it is not ordinary for them yet. The gap was never the tool. It was that the procedures in my head had never been written down for anybody else to run.
I built a free Business Brain AI Audit for this. You give it your email, the audit arrives as one HTML file, and you attach it inside the AI workspace where your business knowledge already lives. Then you paste one line of instruction and send it. Your AI fills in the report and hands it back scored, across the six areas it checks, with the three gaps that matter most named at the end. It costs nothing.
One limit worth knowing before you run it. It reads only what your AI can find inside that one workspace during that one run. It cannot see your other chats, your account memory or your connected tools.
If the audit says the gap is that nothing about your business is written down, that is what Business Brain builds. It is a seven-day course that gives you a knowledge base for AI. That means the plain files Claude and ChatGPT read before they write anything for you. Four modules, about an hour a day, €67 one time, in English, no code and no subscription.
Agent skills, in short
- An agent skill is a folder containing a SKILL.md instruction file that an AI agent discovers and loads on its own when a task matches its description.
- The required contents of a SKILL.md file are a name and a description, followed by the instructions themselves.
- Anthropic introduced the Agent Skills format in October 2025 and published it as an open standard, per an update on its own engineering post dated 18 December 2025.
- The specification now lives at agentskills.io, which calls the format “a lightweight, open format for extending AI agent capabilities with specialized knowledge and workflows.”
- The loading mechanism is called progressive disclosure and runs in three stages, discovery, activation and execution.
- The agentskills.io client showcase listed forty-six agent products on 24 October 2026, among them Claude, ChatGPT, Codex, Claude Code, Cursor, GitHub Copilot, VS Code and Gemini CLI.
- A skill tells an AI how to do something, while MCP connects an AI to a live system it can reach.
- Building a skill requires no code, because a SKILL.md file is plain markdown text.
- This is a software format and is unrelated to the professional abilities of a real estate or insurance agent, and unrelated to the Agent Skills mechanic in the video game Digimon Story: Time Stranger.
- Business Brain is a seven-day course in four modules that builds the written knowledge base those skills draw on, €67 one time, in English, no code required.
- The free entry point is the Business Brain AI Audit at https://hey-julia.kit.com/business-brain-audit.
Frequently asked questions
What are agent skills in AI?
Agent skills are folders of written instructions that an AI agent finds and opens by itself when a task matches what the folder is for. Each one contains a file called SKILL.md with a name, a short description and the procedure itself. The AI reads only the descriptions until something is relevant, then opens the full file.
What is the difference between agent skills and MCP?
A skill tells your AI how to do something, and MCP connects your AI to a system it can reach, such as your calendar or your inbox. Skills are written procedure, and MCP is a live connection. Most solo business owners get more out of skills first, because the undocumented procedure is usually the bigger gap.
Can ChatGPT use agent skills?
Yes. The format started at Anthropic and was published as an open standard, and the agentskills.io client showcase lists ChatGPT and Codex alongside Claude, Claude Code, Cursor, GitHub Copilot, VS Code and Gemini CLI. The same folder works across supporting tools, so a skill you write once is not locked to one subscription. Check your own tool’s current documentation before you rely on it.
Where do you store them?
Inside the AI tool itself, in the place that tool sets aside for them. For Claude that is the skills area of your account or project, and ChatGPT and the other supporting tools use a skills folder in the workspace. Check your own tool’s documentation for the exact location, since it moves as products update. The file itself is ordinary text, so keeping a copy in your own folders is sensible.
Will they replace MCP?
No, because they solve different problems. MCP exists so an AI can reach a live system and act on it, and no amount of written instruction replaces a connection. The two are designed to sit next to each other, with the skill describing the procedure and MCP providing the reach.
Can one skill reference another?
Yes, and a skill can also point to other files bundled beside it, which is the usual way to keep one file readable. A content skill can reference a separate examples file rather than listing twenty examples in the body. Start with one file and split it only when it stops being easy to read.
When should you write one?
Write one whenever you catch yourself explaining the same procedure to your AI for the third time. That repetition is the signal, and the fix is to write it once into a file rather than retyping it from memory in every new chat. Anything you do the same way each week is a candidate.
Can you train Claude or ChatGPT on your own business with skills?
You cannot retrain the underlying model, which belongs to Anthropic and OpenAI, and you do not need to. What you can do is write your procedures and your business facts into files the AI reads before it answers, which is what people actually want when they ask this. The effect on your output is what most people mean by train, and it beats any prompt trick.
How long does it take to write a skill?
The folder and file take about two minutes, and knowing what belongs inside takes considerably longer. Writing down how you actually work is the real task, and it is the step most people skip. The full set, which is what a knowledge base for AI actually is, takes roughly seven days at about an hour a day if you follow a structure rather than inventing one.
