Skills: Transforming your experts' know-how into reusable assets
In any organization, a crucial part of the value lies in know-how that isn't documented: how your best lawyer reviews a contract, the process your sales director follows to qualify an opportunity, the checks your management controller performs before approving a report. These methods reside in a few key minds. When those minds leave, the methods disappear with them. "Skills," a recent feature of AI assistants, offer for the first time a simple way to capture this know-how and make it accessible to everyone.
What is a skill?
A skill is a set of instructions that AI automatically loads when the task is suitable. In practical terms, it's a small file containing instructions written in natural language: the method to follow, the quality criteria, known pitfalls, the templates to adhere to, and possibly examples of good and bad deliverables.
The difference with a reusable prompt is fundamental. A prompt requires you to remember to copy and paste it at the right moment. A skill triggers automatically: the user asks "review this supplier contract," and the assistant applies the company's review method, without anyone having to remind them. The skill ceases to be an option and becomes the default behavior.
From individual expertise to collective assets
The true potential of skills is not individual, it's organizational. Let's take an example we often encounter on assignment: a sales team where each member writes their proposals in their own way. The best proposals implicitly follow a proven structure: restating the need, a three-option approach, supporting evidence through references, and clear conditions. An expert can dictate this structure in a one-hour meeting.
Transformed into a shared skill, it becomes an asset: every team member, from intern to senior, now produces proposals that adhere to the company's methodology. Juniors learn the methodology by observing its application. Seniors save time on formatting and dedicate it to substance. And when the methodology evolves, the skill is modified once, and the entire team benefits immediately.
This represents a fundamental shift in knowledge management. Traditional document repositories store knowledge that no one ever revisits. A skill, on the other hand, is knowledge that is put into practice.
How to write a good skill
Three principles emerge from the practice.
Start with an expert, not a blank page. The best skill is the transcription of real-world expertise. The most effective method: ask the expert to perform the task while commenting aloud, record it, and then have the AI structure the transcription. A first, accurate skill is obtained in half a day, which the expert then refines.
Write criteria, not generalities. “Write a good summary” teaches the model nothing. “The summary should be one page long, begin with the requested decision, quantify each impact, and explicitly indicate points of uncertainty”: these are instructions that change the outcome. Prohibitions are just as important as guidelines: something we never want to see in the deliverable.
Test on real-world cases before releasing. A skill is validated like a procedure: it's tested against five or six past cases where the correct outcome is known, compared, and adjusted. This short cycle prevents the dissemination of a method that seems correct on paper but deviates in practice.
Your first skill, step by step
To get started, a modest but actually used skill is better than a vast, never-consulted one. A good first target meets three criteria: a frequent task, at least weekly; a stable method that the team already agrees on; and a result that can be verified in a few minutes. Meeting minutes, responding to complaints, qualifying a lead, or proofreading a template document are perfect candidates. Strategic negotiation is not.
In terms of structure, a five-section skeleton is sufficient, and it can be reused for all subsequent skills:
- When to use this skill : the situations that trigger it, and those that do not concern it;
- The method : the steps in the order the expert actually follows them, not in the order of the manual;
- Quality criteria : what distinguishes an accepted deliverable from a rejected one, with supporting figures;
- The prohibitions : formulations, commitments or shortcuts that should never be produced;
- An example : a good, real, anonymized deliverable is worth ten pages of explanation.
In practice, this feature is called "skills" at Claude; a similar result is achieved with the instructions of a ChatGPT agent at OpenAI or a Gem at Google. The writing logic is the same everywhere, and that's what matters: a methodology document written to be executed, not archived. Allow half a day for the first version, including the interview with the expert, then an hour of testing on five completed cases before any distribution to the team.
Governance, a prerequisite for trust
A shared skill commits the organization: if the encoded method is flawed, it will be poorly applied on a large scale and consistently. Hence, three simple rules. Each skill has an identified owner, responsible for its accuracy. Skills that touch on regulated areas (legal, HR, finance) are validated by the relevant function before distribution. And versions are date-coded, because a work method evolves like a quality document.
This governance is anything but cumbersome bureaucracy: it's exactly what certified organizations already know how to do with their procedures. The novelty is that the procedure is no longer gathering dust in a filing cabinet; it's actively used.
An asset that is increasing in value
AI models are replaced every six months; skills, however, remain. Written in natural language, they survive tool changes and are enriched with use. Organizations that begin today to capture their critical know-how are building an asset that their competitors cannot copy: their AI models will be the same, but their methods will not. In the race for AI, models are commodities. What you teach them is not.
To learn more: read our analysis of context engineering and MCP, Enterprise AI USB-C Port.