A skill is a method written once, which the AI then follows with every request. The idea is simple and its value shows quickly. Yet we realised that today a number of users rely on skills, but not enough of them.
A skill tied to a single AI
A skill is most often written in one particular AI, by one particular person. It stays there. For a colleague to find it, someone has to send it. For it to work in another AI, it has to be rewritten. And when a firm has several AIs, which is becoming common, each team redoes its own work.
We drew two ideas from this. There was something to do about sharing skills, and the skill needed to be detached from the AI.
Share
The skills worth having are the ones an expert wrote or reviewed: contract review as the firm practises it, the committee note as the legal department wants it. They should be in one place, available to the whole team, rather than in someone’s history.
Detach
The skill does not belong to an AI. It lives separately and plugs into the one you use, thanks to the MCP protocol. It can be moved from one AI to another without being rewritten. And when a firm has several AIs, the same skill works in each of them.
What a skill contains
- the steps, in the order the expert follows them;
- the checkpoints that must never be skipped;
- the sources to consult, and in what order;
- the firm’s reporting format;
- the cases where the AI must stop and ask.
It is written with the expert, reviewed by them, and can be changed at any time. It stays readable, so it can be checked.
Three skills in service
- Contract review. “Review this supplier contract using our method.” Price, liability, termination, in the expert’s order. Delivered as an internal note, gaps first.
- Prior-rights search. “Is this brand name available?” The databases to query, the similarity criteria, the summary in the firm’s format.
- Committee note. “Prepare the note for Thursday’s committee.” The answer in the first line, three options at most, one recommendation.
What it changes
The know-how stays with you. There is a single version to keep up to date, by the expert who wrote it. You change AI without rewriting anything, and the whole team works with the same method, whichever AI they open.
For more on the module, see the Skillshub page.