The most important AI standard of 2025-2026 is one almost nobody is calling a standard. Agent Skills, the SKILL.md format Anthropic introduced for Claude Code in late 2025, now runs in nine AI tools across five vendors. It has become the portable unit of AI workflow. And almost no commentary has noticed.
This article is the long-form treatment. What the standard is, where it came from, the nine platforms that support it today, the practical implications for how you write AI workflows, and where the standard is heading.
If you read Cowork plugins explained for SharePoint AI Skills practitioners first, you have the applied context. This article zooms out to the format itself.
What Agent Skills actually is
Strip away the marketing and a Skill is simple: a Markdown file with YAML frontmatter, named SKILL.md.
---
name: meeting-prep
description: Use when user asks to prepare for an upcoming meeting. Pulls calendar context, recent email threads, and relevant documents into a briefing document.
---
You are preparing a meeting briefing for the user.
## Steps
1. Identify the meeting from the user's request (calendar entry or topic).
2. Pull the most recent five email threads with attendees.
3. Find any documents linked from the calendar invite or referenced in the threads.
4. Produce a one-page briefing with sections: meeting purpose, attendees, key context, suggested talking points.
## Output format
A Word document titled "Briefing: [meeting topic]" saved to the user's OneDrive in the Briefings folder.
That is it. Two required frontmatter fields (name, description), instructions in the Markdown body, optional reference materials in companion files alongside the SKILL.md.
The format is deliberately minimal. It does not specify a runtime. It does not require an SDK. It does not depend on a particular vendor's API. Anyone with a text editor can author one.
The minimalism is the point. By keeping the format slim and open, the same file becomes a portable artefact that runs anywhere a Skills-compatible runtime exists.
The nine platforms that support it today
Microsoft's Cowork plugin documentation lists nine platforms as fully compatible with the Agent Skills standard.
| Platform | Vendor | Notes |
|---|---|---|
| Microsoft 365 Copilot Cowork | Microsoft | Used as plugin skills and as user-level custom skills in OneDrive |
| AI in SharePoint | Microsoft | Site-scoped skills at /Agent Assets/Skills/<name>/SKILL.md |
| Claude Code | Anthropic | Origin of the format. Plugin-level and project-level skills |
| Claude.ai Projects | Anthropic | Skills uploadable as project files |
| Visual Studio Code Copilot | Microsoft | Agent Skills supported in agent mode |
| Gemini CLI | Agent Skills supported | |
| JetBrains Junie | JetBrains | Agent Skills supported |
| OpenAI Codex | OpenAI | Agent Skills supported |
| Cursor | Anysphere | Agent Skills supported |
Five vendors. Nine products. One file format.
This is broader adoption than most industry-standard AI formats achieve. Compare to the Model Context Protocol (MCP), another open standard that has become widely adopted. Or to function calling formats, where each vendor has its own conventions despite years of convergence pressure. Agent Skills has converged faster than either.
The reason is partly that the format is simple. Markdown plus YAML is universal. The barrier to adoption is essentially zero for any vendor that wants to support reusable workflows.
The other reason is that the format separates intent from runtime. A SKILL.md describes what the workflow should do. The runtime decides what the workflow can actually do. Vendors can adopt the format without committing to compatibility on capabilities.
Where it came from
Anthropic introduced the Skills concept in late 2025 as part of Claude Code's plugin architecture. The original use case was developer productivity: a structured way to give Claude Code repeatable instructions for things like running tests, formatting code, or generating boilerplate.
Microsoft adopted the format for AI in SharePoint when the Knowledge Agent shipped in preview (later renamed AI in SharePoint). Site owners author SKILL.md files at /Agent Assets/Skills/<name>/SKILL.md and they get loaded into Cowork-style runtime in the SharePoint context.
Cowork itself launched in March 2026 with built-in skills using the same format, plus support for user-authored custom skills in OneDrive. The May 2026 plugin announcement added the packaging layer (M365 App Packages wrapping SKILL.md files for tenant-level distribution).
Google, OpenAI, JetBrains, and Cursor each adopted the format for their own agent products through 2025 and early 2026. None of them announced "we are adopting Anthropic's standard" with fanfare. The format simply showed up in their docs as the way to write reusable workflows.
The pattern is similar to how Markdown itself became a standard. No formal committee. No vendor-driven blessing. The format was simple enough and useful enough that adoption became the path of least resistance.
What portability actually means
The promise of a cross-platform standard is "write once, run anywhere". The reality of Agent Skills is "write once, adapt lightly per target".
A SKILL.md file is parseable in any of the nine platforms. The frontmatter and the body are read the same way. The format itself is truly portable.
What is not always portable: the steps inside the body. A SKILL.md that calls an external API works in Claude Code (which can run shell commands and HTTP requests) but not in AI in SharePoint (which has no external system access). A SKILL.md that creates a Word document works in Cowork (which has the Word skill loaded) but does not work in Cursor (which is a code editor).
The practical pattern: write the SKILL.md for the most capable target, then maintain per-target variants where the runtime constrains the steps.
| Target | Runtime capabilities | Best fit for |
|---|---|---|
| Claude Code | Shell, HTTP, file system, MCP servers | Developer workflows, build scripts, repository operations |
| Cowork | M365 surface (Outlook, Teams, Word, Excel, OneDrive, calendar, search) plus connector plugins | Knowledge work spanning multiple M365 apps |
| AI in SharePoint | One SharePoint site's content (libraries, lists, pages) | Site-scoped, team-shared workflows |
| Copilot Studio agent | Cross-system via connectors, Power Automate flows | Cross-platform, action-taking, multi-channel |
| VS Code Copilot | Editor file system, integrated terminal | Editor-bound workflows, code generation |
| Gemini CLI | Google ecosystem plus shell | Google Workspace and CLI workflows |
The same workflow intent (e.g. "produce a weekly project report") gets implemented differently per target. The SKILL.md in each version describes the same goal; the steps are written for the runtime.
For most authoring teams, the right pattern is one canonical SKILL.md per workflow plus per-target editing. Three SKILL.md files for three runtimes is fine. A 30-line abstraction layer is not. Skills are meant to be readable; over-engineering them defeats the purpose.
What this means for SharePoint AI Skills practitioners
Three implications worth naming.
One: every SharePoint AI Skill is potentially nine-platform deployable. The Markdown file you have already written is in the standard format. Where it deploys depends on the work, not on the file. A site-scoped policy lookup stays a SharePoint AI Skill. A workflow that spans calendar, email, and documents becomes a Cowork plugin. A developer-facing version of the same intent could become a Claude Code plugin.
Two: the investment in writing good SKILL.md files compounds. Time spent making a workflow clear, well-described, and well-structured pays back across multiple platforms rather than locking you into one vendor's tooling. This is a different return profile from investing in a vendor-specific equivalent (like a Power Automate flow or a Copilot Studio topic). The vendor-specific options are deeper but locked in. SKILL.md is shallower but portable.
Three: the authoring discipline matters more than the platform choice. What makes a SKILL.md work well is the same on every platform: a clear description that includes trigger phrases, instructions that read like a runbook, examples that show the desired output, edge case handling. Get the authoring discipline right and your skill works well everywhere. Get the description vague and the skill is unreliable on every platform.
The companion how-to is in how to ship your SharePoint AI Skill as a Cowork plugin. The runtime comparison across surfaces is in Cowork Skills and SharePoint AI Skills.
What Microsoft's adoption tells you about the trajectory
Microsoft's commitment to Agent Skills is now structural, not experimental.
Three signals.
Cowork built on the standard. Cowork's 13 built-in skills are SKILL.md files. Custom user skills are SKILL.md files. Plugin skills are SKILL.md files. The standard is the foundation of Cowork's architecture, not an optional extension.
AI in SharePoint built on the standard. When Microsoft renamed Knowledge Agent to AI in SharePoint and rebuilt the runtime on Anthropic's Claude, the SKILL.md format stayed. The site-scoped Skill format that SharePoint AI Skills use is the same Agent Skills format.
The conversion script ships. Microsoft built and documented Convert-ClaudePluginToMOS3.ps1 specifically to take Claude Code plugins (which use the standard) and turn them into M365 App Packages. The script's existence signals that Microsoft expects developers to author in the open standard first and target M365 second.
For an enterprise vendor, that level of commitment to a standard originally introduced by a competitor is unusual. It signals confidence that the format is durable enough to bet on.
What the limits of the standard look like
Three limits worth knowing before you over-invest in portability.
Runtime capability differences are real. A skill that uses Power BI in Cowork does not work in Claude Code without the Power BI MCP server. A skill that organises files on disk in Claude Code does not work in AI in SharePoint. Plan for per-target adaptation; do not assume zero editing.
Companion file conventions vary. All nine platforms accept reference material alongside SKILL.md, but the supported file types and size limits differ. Microsoft's Cowork allows 20 companion files, 5 MB each, 10 MB total per skill. Claude Code allows different limits. Build skills with light dependencies and they port more easily.
Vendor-specific extensions are starting to appear. Microsoft has added cowork.category and cowork.icon to its SKILL.md frontmatter. These are valid Agent Skills frontmatter (the standard allows arbitrary fields) but they only do anything in Cowork. Other vendors will likely add their own extensions. The base format stays portable; specific extensions are platform-specific.
The pattern looks like Markdown's evolution. A stable core that runs everywhere, plus optional vendor extensions that activate features on specific platforms. The discipline is to write the core SKILL.md cleanly and use extensions sparingly.
What good authoring looks like in 2026
Three principles that hold across all nine platforms.
Write descriptions like trigger phrases. "Use when the user asks for a meeting briefing" beats "Helps with meetings". The description is what the runtime uses to decide whether to load your skill. Specific is better than aspirational.
Keep SKILL.md focused, push details to companion files. A 1,500-2,000 word SKILL.md is the sweet spot. Longer skills are harder for the runtime to load and harder for humans to maintain. Move detailed reference content to a references/ subfolder.
Test on the runtime, not in the abstract. A SKILL.md that reads well does not always perform well. Run it on the target platform with realistic inputs and edit based on actual outputs. Five rounds of test-and-edit beats five rounds of authoring without testing.
These principles do not depend on which platform you are targeting. They are properties of well-authored Skills, full stop.
Where the standard is heading
Three predictions worth sharing, with the caveat that all predictions about AI in 2026 should be held loosely.
Wider vendor adoption. Several smaller AI tools have not yet adopted Agent Skills. Most will, because the cost of supporting it is near zero and the benefit (their users can bring existing skills) is real.
Tighter integration with MCP. Skills describe intent. MCP describes capability. The two are increasingly bundled in plugin packages (Microsoft's Cowork plugin format is the clearest example). Expect more vendors to ship combined Skills + MCP packaging.
More vendor-specific frontmatter extensions. Microsoft has started; others will follow. The base format will stay stable. The optional extensions will activate platform-specific features. As long as the extensions are optional, portability survives.
The longer-term trajectory is that Skills become the lingua franca for AI workflow description across the industry. Not because anyone declared it the standard, but because the format works, adoption is broad, and the economics favour authoring once over authoring per vendor.
For SharePoint AI Skills practitioners, the implication is simple: you are already writing in the format that runs everywhere. Keep writing. Pick the platform per workload. Let the portability work in your favour.