Every few months, someone announces a new AI platform. A new agent framework. A new way to build "AI-powered" anything. And every time, the same question comes up: do we rebuild everything?
The answer, increasingly, is no. Because the smartest organisations aren't building for platforms. They're building skills.
What Is a Skill?
A skill is a modular, reusable capability that an AI agent can draw on to do a specific job. Draft an RFP response. Triage a support ticket. Analyse a spreadsheet. Generate a branded email. Process a claim.
If an agent is a worker, a skill is something that worker knows how to do. And just like human skills, AI skills can be taught, refined, documented, and — here's the important part — transferred.
A well-built skill isn't locked to one platform. It's a structured definition of what to do, how to do it, what to watch out for, and what good looks like. That definition can be loaded into Claude Code, Microsoft Copilot Cowork, ChatGPT, Codex, Gemini, or whatever comes next.
The skill is the asset. The platform is just where it runs.
Why Skills Matter More Than Agents
There's a lot of excitement about AI agents right now. And agents are genuinely powerful — they plan, execute, iterate, and produce real outputs autonomously. But here's what most conversations miss:
An agent is only as good as the skills it has access to.
You can give an agent the most powerful model in the world. Without the right skills — the domain knowledge, the process logic, the brand voice, the compliance guardrails — it's just a very expensive generalist that doesn't know your business.
Skills are where your organisation's expertise becomes machine-usable. They're the bridge between "AI can do things" and "AI can do our things."
The Portability Breakthrough
Until recently, building for one AI platform meant rebuilding for another. A ChatGPT Custom GPT couldn't run in Claude. A Copilot agent couldn't transfer to Gemini. Every platform was its own island.
That's changing fast:
- AGENTS.md — adopted by over 60,000 open-source projects — provides a standard way to define agent behaviour that works across platforms
- Skill files designed for Claude Code now work in Codex CLI, Gemini CLI, Cursor, and GitHub Copilot
- MCP (Model Context Protocol) standardises how agents connect to external tools and data — plug-and-play across platforms
- A2A (Agent-to-Agent Protocol) enables agents on different platforms to communicate and collaborate
- ServiceNow opened its entire agent skills platform to every developer, from any tool
- The Agentic AI Foundation (Linux Foundation + OpenAI) is building open standards to prevent ecosystem fragmentation
The practical implication: build once, run anywhere. A skill you build today isn't trapped in one vendor's ecosystem. It's an asset that appreciates as the platform landscape evolves.
From GPTs to Coworkers: The Conversion Path
Here's something most organisations haven't realised yet: the skills they're already building can be converted across formats.
A Custom GPT you built in ChatGPT? The instructions, the knowledge base, the behaviour rules — that's a skill definition. It can be restructured into a Claude skill, a Copilot agent, or a Codex command.
A Copilot agent you configured in Microsoft 365? The prompts, the data connections, the output rules — that's portable expertise. It can become a skill that runs in any harness.
A Claude Code slash command that handles your email triage? That's a skill file with trigger logic, step-by-step instructions, and guardrails. It can be adapted to work inside Copilot Cowork, where it runs autonomously across your entire M365 environment.
The conversion isn't automatic — each platform has its own format and capabilities. But the thinking transfers. The domain knowledge, the process logic, the edge case handling — that's the hard part, and you've already done it.
Skill Stacking: Where It Gets Powerful
Individual skills are useful. Stacked skills are transformational.
Skill stacking is the practice of combining multiple skills into compound workflows that handle complex, multi-step business processes. Instead of one agent doing one thing, you orchestrate a sequence of skills that together accomplish something none of them could do alone.
Here's what that looks like in practice:
Example: New Client Onboarding
Instead of building one massive "onboarding agent," you stack skills:
- Intake Skill — extracts key information from the client application
- Risk Assessment Skill — analyses the client profile against your underwriting criteria
- Document Generation Skill — produces the welcome package, policy summaries, and compliance documents
- Email Drafting Skill — composes personalised onboarding emails in your brand voice
- CRM Update Skill — logs everything in your system of record
- Calendar Skill — schedules the first check-in call
Each skill is built, tested, and maintained independently. Each one can be reused in other workflows. The intake skill also works in your renewal process. The email drafting skill also works in your sales outreach. The document generation skill also works for your RFP responses.
The more skills you build, the more workflows you can assemble. This is the compounding return on skill investment — every new skill multiplies the possible combinations.
Example: Weekly Intelligence Briefing
Stack these skills and run them on a schedule:
- Research Skill — scans industry sources for relevant developments
- Analysis Skill — identifies patterns, risks, and opportunities
- Summarisation Skill — distills findings into an executive briefing
- Distribution Skill — formats and delivers to the right people
Built once, runs every week, gets better as you refine each skill. No human assembles the briefing — but a human reviews it before it goes out.
Building a Skills Library
The organisations that will have the strongest AI capabilities in 2027 are the ones building their skills library right now. Here's how to think about it:
1. Audit Your Repeatable Work
Every process your team does more than once is a candidate for a skill. Start with the ones that are well-defined, have clear inputs and outputs, and where quality can be verified.
2. Document Before You Automate
A skill starts as documentation. Before any AI touches it, write down: what triggers the work, what information is needed, what steps are taken, what the output looks like, what the exceptions are, and how quality is verified. This documentation is the skill — the AI execution layer comes after.
3. Build Modular, Not Monolithic
Resist the temptation to build one agent that does everything. Build small, focused skills that do one thing well. Then compose them. A monolithic agent is fragile and hard to maintain. A library of skills is resilient and endlessly recombinable.
4. Version Control Everything
Skills evolve. Your brand voice changes. Your compliance requirements update. Your process improves. Every skill should be version-controlled so you can track what changed, when, and why. This is especially critical in regulated industries where audit trails matter.
5. Test Across Platforms
Don't build exclusively for one AI platform. Test your skills in at least two environments. This keeps them portable and ensures you're building the knowledge, not just the platform integration.
The Enterprise Skills Gap
Here's the uncomfortable truth: most organisations have zero documented AI skills. They have people who know how to prompt ChatGPT. They have a few custom GPTs that one person built. They might have a Copilot agent or two.
But a governed, documented, version-controlled library of enterprise skills that captures institutional knowledge and runs across platforms? Almost nobody has that yet.
That's the gap. And it's a massive opportunity.
The organisations that close this gap first will have a compounding advantage: more skills means more workflows, more workflows means more capacity, more capacity means faster execution. Their competitors will still be asking "which AI tool should we buy?" while they're already running stacked skill workflows that handle work that used to take teams of people.
Where to Start
- Pick three workflows your team does every week. Document them as skills — triggers, inputs, steps, outputs, exceptions.
- Build the first skill in whatever platform your team already uses. Get it working, get feedback, refine it.
- Port it to a second platform to prove portability. This forces you to separate the domain knowledge from the platform-specific implementation.
- Stack two skills together into a compound workflow. Experience the multiplier effect firsthand.
- Start the library. Create a shared repository where skills are documented, versioned, and discoverable by anyone on the team.
The models will keep getting better. The platforms will keep changing. But the skills your organisation builds — the codified expertise, the documented processes, the institutional knowledge made machine-readable — that's yours. It transfers. It compounds. And it's the thing that will actually differentiate your organisation in a world where everyone has access to the same AI.
Build the skills. Stack the skills. Own the skills.