Agent Skills: write down how your company works
Skills turn a procedure, its templates and its rules into a folder an AI loads when needed. Why small teams should care, and how to write one that gets used.
Every business has a way of doing things that lives in people's heads. How a quotation is laid out. Which discounts need the owner's approval. What goes in the monthly report and in what order. How we reply to an unhappy customer. New staff learn it over months, mostly by getting it wrong and being corrected.
AI tools have the same problem, only worse. Each conversation starts from nothing. You can paste in your instructions every time, and most people who use AI at work end up with a document of prompts they copy from. It works, but it is fragile, and it does not scale past one person.
In October, Anthropic released a feature called Skills that tackles this directly, and in December it published the format as an open standard so other AI tools can use the same files. It was one of the least flashy AI releases of 2025 and, for a small team, one of the most useful.
What a skill is
A skill is a folder. Inside is a file called SKILL.md with a short description at the top and instructions underneath, written in plain language. The folder can also hold anything the task needs: a template, a sample of good work, a price list, a script that does a calculation.
The clever part is how it gets loaded. Claude reads only the short description of each skill at first, a line or two. When a task matches, say someone asks for a quotation, it pulls in the full instructions and files for that skill alone. So you can have dozens of skills without stuffing all of them into every conversation.
Anthropic ships some of its own, including skills for producing proper Excel, Word and PowerPoint files. The interesting ones are the ones you write.
Skills worth writing first
The best candidates are tasks that happen often, follow a pattern, and currently depend on one person remembering how it is done.
- Quotations and proposals in your format, with your standard terms, your payment schedule and the questions to ask before pricing.
- The monthly or weekly report: which numbers, from which exports, compared against what, written for which reader.
- Customer replies for the ten situations your team handles most, with the policy behind each and the point at which a person takes over.
- Your brand voice: words you use and avoid, how formal to be in English and in Arabic, how you sign off.
- Checking documents before they go out: invoices against the purchase order, contracts against your standard clauses.
How to write one that gets used
Most of a skill's quality comes from how specific it is. A few rules we follow:
- One job per skill. "Write a quotation" is a skill. "Help with sales" is not.
- Make the description say when to use it, in the words people actually use. If your team says "estimate" rather than "quotation", say both.
- Write the steps the way you would explain them to a capable new hire on their first day, including the judgement calls.
- Include one or two real examples of good output. Models learn the shape of the answer from examples faster than from rules.
- Say what not to do. "Never quote delivery dates" is worth more than a paragraph about being careful.
- Test it on five real tasks before you share it, and fix whatever it gets wrong.
Give every skill an owner. Prices change, policies change, and a skill with last year's price list will produce wrong quotations very consistently.
Why the open standard matters
When a feature belongs to one company, building on it is a bet on that company. In December, Anthropic opened Skills as a standard anyone can implement. The same month, the Linux Foundation set up the Agentic AI Foundation to look after MCP, the protocol that connects AI tools to other software, along with AGENTS.md, a similar plain-text convention from OpenAI for telling coding agents how a project works.
The industry is settling on simple, readable files for telling AI how your organisation works. They live in your own folders and should keep working whichever AI tool your team uses next year.
One caution
Skills can contain scripts, and scripts run. Treat a skill from the internet the way you would treat any software: install it only from a source you trust, and read what it does first. Never put passwords or API keys inside a skill.
The skills you write yourself are a different matter. They are mostly documentation, and writing them has a side benefit people do not expect. To explain a task clearly enough for an AI, you have to explain it clearly, full stop. Often the written-down version is the first time everyone agrees on how the job is supposed to be done.
