The AI advantage is no longer in the model but in a text file that knows how your organisation works.

For two years, competing in AI meant competing to have the best model. Every lab boasted about scoring higher, reasoning better or processing more information than the rest. That race is losing its meaning. Large models are becoming increasingly similar in capability, and when they are all good, the difference is no longer in which one you use. It is in what your organisation builds around it.

And here something unusual is happening. Three labs that, although constantly competing and facing off, Anthropic, OpenAI and Google, have agreed to give the same answer and have adopted the same format to implement it. That answer is called a skill.

An expert without context

Do a test. Ask Claude, ChatGPT or Gemini to review something from your company that they have never seen, for example a business proposal or a campaign brief. The response will sound intelligent. It will be competent, well-written and have no idea how you work, but it won't know what criteria you use to decide, how you construct an offer or who has to sign off before something goes out to the client. 

It is not an intelligence problem. It is like a brilliant consultant on their first day of work. They know an awful lot, but they don't know your business, and they also forget everything you tell them as soon as the conversation ends. That problem cannot be fixed by hiring someone even more brilliant. It is fixed by explaining how you work. That, exactly, is a skill.

From prompts to skills

These last two years have been spent writing better prompts. A good prompt solves a specific task in a specific conversation and vanishes when the session is closed. The next day you go back to explaining to the model who you are and what you expect.

A skill breaks that loop. Anthropic formalised it in October 2025 as a text file, a SKILL.md, with instructions that the agent reads before starting to work. Without code and without an infrastructure team. Their documentation summarises it in an idea that we share. Create once, use always.

The operational effect is consistency. The agent stops improvising your process in each session and shifts to following your method, with your criteria and your quality standard. 

We should be precise. A skill does not guarantee a character-for-character identical result, because models are probabilistic. What it eliminates is the variability of the process, and for an operational process, that is precisely what matters.

Three labs, one conclusion

The striking thing is not that the idea exists, but who endorses it. 

• Anthropic defines skills as reusable resources that turn a generalist agent into a specialist. 

• OpenAI describes them in its documentation as modular instructions to code processes and conventions, from corporate style guides to multi-step workflows. 

• Google highlights in Gemini that they ensure the consistent execution of complex tasks through a procedural framework. 

Three ways of saying the same thing.

Anthropic took the first step. They formalised the concept in October 2025 and in December published the specification as an open standard, which OpenAI and Google adopted in a matter of weeks. 

The format is already reaching the tools that businesses use on a daily basis. Microsoft 365 Copilot, for example, allows you to add your own skills by simply dropping a SKILL.md file into OneDrive. 

A skill is an asset in your business

A GPT's instructions are invisible to everyone except the person who created it, who cannot even export or version them easily. The day that person leaves, the knowledge walks out of the door with them.

It is the same problem businesses had with spreadsheets in the nineties, business information trapped in individual files. The solution was to move it to shared systems with version control and institutional ownership. Now the solution is the same: to extract that knowledge from closed platforms and make it visible, portable and organisational.

And here is the fundamental shift: a skill is a company asset. They are your processes written in an open format that you own and version like any other corporate document.

Why skills beat code

A skill introduces four capabilities that no traditional software module can match:

- Created by the business: they can be defined by non-technical profiles, without the need to program.

- No deployments: changes are applied instantly, without release cycles.

- Readability and auditability: they are easy to understand and review compared to complex coded solutions.

- Flexibility: they can be removed or adjusted according to their value with no structural impact.

- Real delegation: they work like an organizational onboarding for AI agents, transferring context, criteria and ways of working in a repeatable and scalable manner.

Altogether, skills turn organizational knowledge into an operational layer that allows AI to work with the same criteria, context and consistency as the business itself, in a scalable and reusable way.

The connection with agents

At RocaSalvatella we have been arguing that the significant leap in AI is not co-pilots, but agents capable of executing entire processes. This is where skills fit in.

An autonomous agent, if you do not share your methodology with it, will make it up. It produces something that looks correct and that a professional would not sign off on. Skills are the layer that prevents this: the institutional knowledge that makes an agent work within your way of doing things and under your governance. 

The competitive advantage will not be in the model you use, but in how your organisation is able to organise, control and leverage its own knowledge as an asset of its own, without depending on a specific provider.

What we recommend from RocaSalvatella

1. Audit your invisible skills. How many GPTs, GEMs and assistants already exist scattered across the organisation. That is knowledge you do not control. Start extracting it and observe which teams it concentrates around.

2. Choose the first skill. Look for the process where your people repeat the same context in each session. Publish it and iterate it.

3. Make it easy to contribute. The skill improves when people use it and give feedback on it. Make it easy for anyone to say what works and what doesn't, and thank those who do publicly. People will join in because it is useful to them, not because you tell them to.

4. Think in branches. Imagine your AI capacity as a tree: the trunk is the infrastructure, the branches are your skills and the leaves are the sessions, the actual work. Most companies pour their energy into the trunk. The return is in the branches.

The knowledge has always been there. What is new is not the knowledge, it is its capacity for execution. And now, at last, it fits in a text file.