Anthropic Agent Skills: The Open Standard That Competitors Can’t Ignore

Anthropic just gave away another piece of infrastructure that competitors will have to adopt or fight. On December 18, 2025, the company announced Agent Skills as an open standard—ten days after donating the Model Context Protocol (MCP) to the Linux Foundation. The pattern is unmistakable: Anthropic is positioning itself not as a platform owner, but as the architect of how AI agents work across the entire industry.

And it’s working. OpenAI has already quietly adopted the Skills specification in ChatGPT and Codex CLI. Microsoft implemented it in VS Code and GitHub Copilot. The very companies Anthropic competes with are building on its standards.

What Are Agent Skills, and Why Should You Care?

Agent Skills are essentially instruction manuals for AI agents—directories containing a SKILL.md file with metadata and detailed guidance on how to accomplish specific tasks. Unlike traditional prompts that you paste into a chat box, Skills are persistent, reusable, and automatically discoverable.

The architecture is elegantly simple:

skill-name/
├── SKILL.md          # Required: Metadata + instructions
├── scripts/          # Optional: Executable code
├── references/       # Optional: Reference documentation
└── assets/           # Optional: Supporting files

What makes Skills clever is their “progressive disclosure” loading pattern. At startup, an agent only reads the name and one-line description of each installed skill. The full instructions load only when relevant, keeping token usage efficient. It’s the difference between carrying an entire library in your backpack versus knowing where to find the right book when you need it.

The Open Standard Strategy: Why Give Away the Recipe?

Anthropic published the Skills specification at agentskills.io with full documentation and a GitHub repository. The message is explicit: write a skill once, deploy it everywhere—Claude, ChatGPT, Cursor, VS Code, GitHub Copilot, or any platform that adopts the standard.

This follows Anthropic’s December 9 move to donate MCP to the newly-formed Agentic AI Foundation under the Linux Foundation—with OpenAI and Block as co-founders. We covered the foundation launch when it happened. Two open standards in ten days. That’s not a coincidence. It’s a strategy.

The logic is counterintuitive but powerful: if Anthropic sets the standards everyone adopts, they gain influence without the baggage of vendor lock-in. Developers and enterprises increasingly distrust proprietary ecosystems that could trap them with a single provider. By giving away infrastructure, Anthropic builds trust while simultaneously defining how AI agents should work.

MCP vs. Skills: The Two-Layer Stack

If you’ve been following Anthropic’s recent moves, you might wonder how Skills fit alongside MCP. The answer is they’re complementary, not competing.

Layer	MCP	Agent Skills
Purpose	External connectivity	Procedural knowledge
Function	Connects to APIs, databases, services	Teaches how to use those connections
Format	Protocol specification	Markdown files
Complexity	Requires server implementation	Lightweight filesystem-based

Anthropic’s official explanation: “MCP provides secure connectivity to external software and data, while skills provide the procedural knowledge for using those tools effectively.” Think of MCP as the plumbing and Skills as the instruction manual for the plumber.

Enterprise Features: Why IT Admins Are Paying Attention

Alongside the open standard, Anthropic launched enterprise management capabilities that solve a real problem: how do you deploy AI workflows consistently across an organization without losing control?

The new features for Claude Team and Enterprise plans include:

• Central skill provisioning: IT admins can push skills to all users from admin settings
• Default-on deployment: Admin-provisioned skills are enabled by default, with users able to toggle off if needed
• Partner directory: Pre-built skills from Atlassian, Canva, Notion, Figma, Cloudflare, Stripe, Zapier, and more at claude.com/connectors

This matters because enterprise AI adoption has consistently stumbled on the question of governance. Who controls what the AI can do? How do you ensure consistent behavior across hundreds of employees? Skills give IT departments a lever they’ve been missing.

OpenAI’s Quiet Adoption: Reading the Tea Leaves

Here’s the most interesting development that flew under the radar: OpenAI has already implemented a structurally identical architecture in ChatGPT and its new Codex CLI. Same file naming conventions. Same metadata format. Same directory organization.

OpenAI didn’t announce “we’re adopting Anthropic’s Skills standard.” But they didn’t need to. The implementation speaks for itself. When your competitor builds on your open standard, you’ve won a different kind of market victory—one that doesn’t show up in benchmark comparisons but shapes how the entire industry develops.

This mirrors what happened with MCP. Anthropic released it, and within weeks, Microsoft, Google, and a dozen coding tools had integrated it. Standards have a gravitational pull that proprietary systems lack.

What This Means for Developers

If you’re building AI workflows, the practical implications are significant. Skills are portable. You can develop a skill for your internal processes, test it on Claude, and know it will work on ChatGPT or Cursor without modification. That’s not hypothetical. It’s the explicit design goal.

Creating a skill is straightforward:

---
name: "Code Review Workflow"
description: "Guides thorough code review with security and performance focus"
---

# Instructions

## When to Use
Activate when the user asks for code review, PR feedback, or security analysis.

## Process
1. Check for common vulnerabilities (OWASP Top 10)
2. Evaluate performance implications
3. Assess maintainability and readability
4. Provide specific, actionable feedback

Drop that in a .claude/skills/ directory, and Claude Code will automatically discover and use it when relevant. The same file works in VS Code with Copilot, in Cursor, or on any platform that’s adopted the standard.

The Bigger Picture: Who Sets the Rules for AI Agents?

Zoom out, and the pattern becomes clear. Anthropic has contributed two foundational standards in ten days—MCP for connectivity and Skills for capability. Both are governed by the Agentic AI Foundation, which now includes Anthropic, OpenAI, and Block as co-founders, with AWS, Google, and Microsoft as platinum members.

The implications are substantial. The era of AI development where each company builds proprietary, incompatible systems is ending. We’re moving toward a world where the basic infrastructure—how agents connect to tools, how they learn specialized knowledge—is standardized and interoperable.

Anthropic isn’t just participating in this shift. They’re leading it. And by giving away the infrastructure, they’ve made it very difficult for competitors to ignore. You can build your own proprietary system that does the same thing. But why would you, when the open standard already works and your users expect portability?

That’s the genius of open standards. They constrain competitors not through lock-in, but through adoption. Once the industry converges on a format, going proprietary becomes a competitive disadvantage.

What Comes Next

Expect rapid iteration on the Skills specification. The GitHub repository already has over 20,000 stars, and community-built skills are proliferating. Anthropic’s engineering blog emphasizes that skills should stay under 500 lines and 5,000 tokens—practical constraints that suggest they’re optimizing for real-world deployment, not theoretical capability.

For enterprises considering AI adoption, the message is clear: the standardization wave is here. Skills you build today will work across platforms tomorrow. The investments you make in AI workflows aren’t locked to a single vendor.

And for AI companies watching Anthropic’s strategy, the lesson is equally clear: sometimes the way to win isn’t to build the highest walls around your platform. It’s to tear them down for everyone—including yourself.