Claude Skills for a company: who keeps the rules current?

Claude Skills are one of the more useful things to happen to AI at work this year. You write down how a task is done, once, in a SKILL.md file. Claude loads it when a request fits and follows the steps. Writing an offer, answering a complaint, checking an invoice: things people used to explain in every chat.

For one person, that’s the whole story. For a company, it’s where a new problem starts.

What a skill actually contains

Open a few skills that a team has written and you’ll find two kinds of content mixed together.

The first is the how. “Read the complaint. Check the order. Draft a reply in our tone. Offer a fix before you offer money.” That part rarely changes.

The second is the what applies. “Refunds up to 500 euros without approval.” “Return window 30 days.” “Never promise a delivery date for custom orders.” Those are business rules, and they change. Prices go up, a threshold moves, a lawyer asks for different wording.

As long as both live in the same file, every rule change means editing every skill that mentions it.

Where it breaks

Say your support team has a skill for complaints, your sales team one for offers, and finance one for credit notes. All three mention the refund limit. In March it goes from 500 to 300 euros.

Someone updates the support skill. Sales hears about it two weeks later. The finance skill sits in a colleague’s personal Claude account and nobody remembers it exists. Now Claude gives three different answers depending on who asks, and every one of them looks confident.

On Claude Team and Enterprise, an Owner can make skills available to the whole organisation, which helps with distribution. It doesn’t tell you which rule is current, or who decided it. And if half the company uses ChatGPT or Copilot, they have their own copies anyway.

Split the steps from the rules

The fix is boring and it works: keep the how in the skill and move the what applies somewhere central.

The skill says “check the refund limit for this customer” instead of “the refund limit is 500 euros”. The limit itself lives in one place, written once, with a reason and the name of the person responsible. When the AI runs the skill, it looks the rule up.

Two things make this hold up over time.

Every rule has an owner. The head of finance owns the refund limit. Anyone can suggest a change, but it only counts once finance approves it. That sounds like bureaucracy. In practice it’s one click, and it’s the only way to know which version is the real one.

The AI reads the rule live. Copying rules into each skill, project or custom instruction just recreates the problem. The AI needs to fetch the current version at the moment it works. Today the standard way to do that is MCP (Model Context Protocol), which Claude, ChatGPT and Copilot all support in their business plans.

A quick check for your skills

Take the skills your team uses most and highlight every sentence that would have to change if a policy changed. Prices, limits, deadlines, exceptions, wording you’re not allowed to use.

If that’s a sentence or two per skill, you’re fine. Keep them in mind when policies change. If it’s half the file, or the same rule shows up in five skills, you’re maintaining the same fact in five places. Sooner or later one of them will be wrong.

Where to start

You don’t need new software for the first step. Pick one skill and pull its rules out into a short list: the statement, the reason, who owns it. Then rewrite the skill so it refers to those rules instead of repeating them.

Doing that for one skill takes an afternoon. Doing it for every skill and every AI tool in the company, while the rules keep changing, is the job of a company brain. With CtxCore Brain you upload a SKILL.md, Brain suggests the business rules it depends on, and Claude gets those rules over MCP in the version that applies today. The skill files stay where they are.