How to turn your organisation's knowledge into AI skills, systematically

Start with a map, not a skill. In each team, ask four questions: who gets asked the most questions, which tasks only one person can do, where results depend on who does the work, and who is likely to leave or retire in the next year or two. That last one matters more than most companies admit: in a 2025 study by APQC, 92% of organisations said they do not consistently capture knowledge from people about to retire.

The result is a short list of processes, each with the expert who knows it best. That list is your backlog.

From the list, choose a process that is done often, that the experts agree on, and whose result you can judge. A quarterly report that three people do three different ways is a poor first choice. A weekly client update with a clear template is a good one.

Then check the data around it. A skill can only be as good as the files it works from, and our data checklist tells you in an afternoon whether the process is ready or needs tidying first.

Do not ask the expert to write a document. Ask them to do the task with an AI tool while someone watches, and note every piece of context they add and every correction they make. Anthropic recommends exactly this: work through the task with the AI and notice what information you keep providing. That information is the skill.

Pay most attention to the exceptions. "Unless the client is in healthcare." "Except at the end of the quarter." "Always check this first." Experts apply these rules without thinking, and they are what makes their work better than everyone else's.

Use the process section of What makes a good AI skill as your interview guide. Check first that the experts agree on the process, then for every step ask where the information comes from, which decisions are made, and how the expert knows the step was done right. Those answers become the backbone of the skill.

Turn the notes into a skill with a clear structure: defined steps, the data source for each step, the decision points, a check after each step and a quality check on the result. The AI can help with the writing, but the structure has to come from you. Skills an AI writes on its own, without that input, tend to do worse than no skill at all.

The same article covers the remaining elements, from examples of good output to leaving out what the AI already knows.

Take three to five past cases where you know the right result, including a messy one, and run the skill on each, several times. Then let two or three colleagues use it on their own work. Testing is where the value is: in SkillsBench, a 2026 research benchmark, skills curated by people raised AI success rates from 34% to 51%, while skills the AI wrote for itself did worse than no skill at all.

Expect to go back to the expert at least once. The tests always surface a rule nobody mentioned.

A skill that sits on the expert's laptop has captured the knowledge but not shared it. Put it in a company library where everyone can find it, give it a named owner, and have the right person approve it before it reaches the team. Then make sure the approved version reaches every AI tool your people use, not only the one the expert prefers.

Knowledge goes stale when nobody looks at it, and skills are no exception. The good news is that a skill is used every day, which makes upkeep part of the work. Support teams have long followed a similar principle: review knowledge while you use it, when you still have the context to improve it.

In practice: let users send feedback and suggested changes to the owner, turn each improvement into a new reviewed version, test again whenever the AI model changes, and review each skill on a schedule that matches how much the company relies on it.

The seven steps, and what each one should leave you with.

Once the first skill works, take the next process from the list. After a few rounds you have something most companies lack: a growing library of the knowledge that runs the business, each piece with an owner, a history and a way to improve. That library is a company asset, and like any asset it needs managing.

Fygment is built for exactly that. It holds the library, with an owner, versions and approvals for every skill, delivers the approved version to each AI tool your team uses, routes feedback and suggested changes to the owner, and runs each skill's test cases before a new version is approved. While someone writes a skill, Fygment points out what it is still missing.

If you want to try the method on your first process, Fygment is free to start.