By Arnau Puig de Quadras, RocaSalvatella

The advantage in AI is no longer the model itself. It comes from the knowledge your organisation can give it.

For the past two years, AI competition has been about building better models. Every lab has tried to outperform the others in reasoning, performance and capabilities. That race is rapidly losing relevance. Leading models are becoming increasingly similar in capability, and when they are all highly capable, differentiation no longer comes from the model you use. It comes from what your organisation builds around it.

What is remarkable is that three companies that compete intensely with one another, Anthropic, OpenAI and Google, have arrived at the same conclusion and adopted the same implementation format. That conclusion is called a skill.

An Expert Without Context

Try a simple experiment. Ask Claude, ChatGPT or Gemini to review something from your organisation that it has never seen before, such as a commercial proposal.

The response will sound intelligent, competent and well written, despite having no understanding of how your organisation actually operates. It will not know the criteria you use to make decisions, how you structure an offer, or who needs to validate a deliverable before it reaches a client.

This is not a problem of intelligence. It is similar to a brilliant consultant on their first day at work. They know a great deal, but they do not yet understand the organisation. Worse still, they forget everything you explain as soon as the conversation ends.

The solution is not to hire someone even more brilliant. The solution is to explain how your organisation works. That is exactly  what a skill does.

From Prompts to Skills

Over the last two years, organisations have invested considerable effort in writing better prompts. A good prompt solves a specific task within a specific conversation, then disappears when the session ends. The next day, the model once again needs to be told who you are and what you expect. A skill breaks that cycle.

Anthropic formalised the concept in October 2025 through a simple text file called SKILL.md, which an agent reads before it begins working. No coding is required. No infrastructure team is needed.

The idea is straightforward: create a capability once and reuse it when needed. Instead of improvising your process in every session, the agent follows your methodology, your decision criteria and your quality standards.

It is worth being precise. A skill does not guarantee identical outputs character by character, because models remain probabilistic systems. What it eliminates is process variability. For operational processes, that is what truly matters.

Three Labs, One Conclusion

The key point is not the idea, but who agrees on it.

• Anthropic defines skills as reusable resources that transform a general-purpose agent into a specialist. 
• OpenAI describes them as modular instructions that encode processes and conventions, from corporate style guides to multi-step workflows.
• Google presents them within Gemini as a mechanism for ensuring consistent execution of complex tasks through a procedural framework. Three different descriptions of the same underlying idea.

Anthropic took the first step, formalising the concept in October 2025 and publishing the specification as an open standard two months later. OpenAI and Google adopted the format within weeks.

The approach is already finding its way into the tools organisations use every day. Microsoft 365 Copilot, for example, allows organisations to deploy their own skills simply by placing a SKILL.md file in OneDrive.

A Skill Is a Corporate Asset

The instructions behind a custom GPT are invisible to everyone except its creator, who often cannot easily export, version or govern them. When that person leaves the organisation, the knowledge leaves with them.

This is remarkably similar to the spreadsheet problem of the 1990s: critical business knowledge trapped in individual files. The solution then was to move information into shared systems with version control and institutional ownership. The solution today is much the same: extract knowledge from closed platforms and make it visible, portable and organisational.

This is the deeper shift. A skill is a corporate asset. It is your processes documented in an open format that your organisation owns, manages and versions like any other strategic document.

Why Skills Outperform Code

Skills introduce capabilities that traditional software modules struggle to match. 

• They can be defined by non-technical professionals without writing code. Changes can be implemented immediately, without release cycles.
• They are transparent, understandable and easy to review, unlike many coded solutions. They can be adjusted or removed according to the value they create, without structural impact.
• Most importantly, they enable genuine delegation. Skills function as organisational onboarding for AI agents, transferring context, decision criteria and ways of working in a repeatable and scalable manner.

Collectively, they transform organisational knowledge into an operational layer that enables AI systems to work with the same context, judgement and consistency as the business itself.

The Connection to Agents

At RocaSalvatella, we see the next stage of AI not in copilots, but in agents that can carry out complete processes. This is where skills become essential.

An autonomous agent that is not given your methodology will invent one of its own. It will produce something that appears correct but that no experienced professional would ultimately endorse.

Skills provide the layer that prevents this. They represent the institutional knowledge that enables an agent to operate within your way of working and under your governance framework.

The future competitive advantage will not depend on which model you choose. It will depend on how effectively your organisation can structure, govern and leverage its own knowledge as a strategic asset, independent of any particular technology provider.

What We Recommend at RocaSalvatella

1. Audit Your Invisible Skills: How many GPTs, Gems and AI assistants already exist across your organisation? That is knowledge you do not control. Start extracting it and identify where it is concentrated.
2. Choose Your First Skill: Find the process where people repeatedly provide the same context in every session. Document it. Publish it. Refine it.
3. Make Contribution Easy: A skill improves through use and feedback. Create simple mechanisms for people to share what works and what does not. Recognise those contributions publicly. People will engage because it makes their work easier, not because they are instructed to do so.
4. Think in Branches: Imagine your AI capability as a tree. The trunk is infrastructure. The branches are your skills. The leaves are the individual sessions where real work happens. Most organisations focus their energy on the trunk. The return lies in the branches.
5. A New Form of Execution: The knowledge itself has always existed. What is new is its ability to be executed. 

For the first time, organisational know-how can be encapsulated, distributed and operationalised through something as simple as a text file.