Stop explaining the same thing to AI every day.
AI can explain accounting standards, the history of law and the difference between two types of contracts. But ask it for a monthly report exactly the way you do it, and it starts guessing. Or someone has to explain it. Again. The fix is simpler than it sounds: write the procedure down once and let AI open it on its own.
AI knows the facts. It doesn't know how things are done at your company.
Language models learned from a huge amount of text. They know the facts. What they lack is the working procedure: in what order things are done, what gets checked and what to watch out for at your particular company.
Without a procedure, AI has only two options. Either someone dictates every step each time, or it makes them up. Neither is why you brought AI in.
With a person
An experienced colleague knows that offers for the public sector have a different structure and that a larger discount needs the managing director's approval. A newcomer doesn't know this until someone tells them or they find it in an internal procedure.
With AI
AI is a permanent newcomer. If the procedure isn't written down anywhere, you explain it with every task. If it is written down as a skill, AI finds it and uses it on its own.
A skill is a working procedure in one file.
Technically, a skill is an ordinary text file in a folder. At the top it has a name and a short description of when to use it. Below that are the steps themselves, the rules and examples of a correct result. Templates, reference material or small programs can be attached.
What matters is how AI works with it. At the start, it loads only the names and descriptions of all skills, something like a binder's table of contents. It opens the full procedure only when a task matches the description. It opens attachments only when it actually needs them. That way AI can have hundreds of procedures available without being overwhelmed by them.
Skills are an open format today. The same file works in Claude and in tools from other companies that have adopted the standard. So the procedure isn't locked in with one vendor.
“An accounting firm has a month-end closing procedure for its clients: what gets checked, in what order, which exceptions are handled by phone. Until now, only two people knew it. As a skill, AI uses it at every closing, and a new colleague learns from it at the same time.”
A skill is not the same as access to data.
Companies often mix these up. AI needs three different things, and each is solved by something else.
| WHAT AI NEEDS | WHAT SOLVES IT | COMPARISON |
|---|---|---|
| Facts from your company | Knowledge base (RAG) | Company library |
| Access to systems | Connection via MCP | Keys to the offices |
| How things are done | Skill | Internal working procedure |
A CRM connection gives AI the keys. It doesn't tell it when and why to go into the CRM or what to do there. That is the skill's job. In practice they are combined: the skill says how to proceed, and the connection gives it the tools.
Five rules to recognize a good skill.
1. The description decides whether the procedure gets used at all. AI chooses a skill by its description. “Makes reports” is too little. “Prepares the monthly report for management, use whenever someone wants a report, a closing or a monthly overview” works. The description can safely be more emphatic. Models tend to skip a skill rather than use it needlessly.
2. The procedure is written by whoever does the work, not by AI. If you let AI write a skill for you, you get general advice it already knows. The value is in your experience. The most valuable parts are the exceptions and traps: every correction you make after AI is a line that belongs in the procedure. Otherwise you'll be correcting it next week too.
3. Less is more. AI knows what an invoice or a contract is. The procedure should contain only what it doesn't know: your rules and exceptions. The recommendation is to keep the main procedure under roughly 500 lines and move the details into attachments.
4. Where there's no room for error, let a program decide. AI may interpret the procedure slightly differently each time it runs. For the wording of an email, that's fine. For adding up the amounts in an offer, it isn't. A step that must always be exact should be done by a small program that AI just runs. A program doesn't guess.
5. Check someone else's skill like software. A skill can contain a program that runs on your computer and has access to files. A security audit of nearly 4,000 public skills found that more than a third had a security flaw. So don't install a skill from the internet blindly.
“A law firm prepares the first statement on every new matter using the same structure. The skill contains the structure, forbidden phrasings and two sample statements. Deadline checks are done by a simple calculation, not by AI. The associate gets a draft in a consistent form and corrects the content, not the format.”
When a skill doesn't make sense.
If you do a task once a year, it's not worth writing it down as a skill. The same goes for a procedure that changes every week or that nobody can describe properly yet. A skill captures what already works. It won't figure out for you what the work should look like.
Ask yourself
Which procedure at your company do you keep explaining over and over to new people and to AI?
