Clarion Street

The Skill Standard: How a Markdown Folder Became the Portable API for Agent Capability

·7 min read

A skill, in the sense that Claude Code, Cursor, Gemini CLI, and a few dozen other agent products now mean it, is a folder. The folder holds a SKILL.md file with a name and a description in YAML frontmatter and Markdown instructions, and it can optionally bundle scripts/, references/, and assets/ directories. That is the whole format. It began as Anthropic’s way to hand Claude reusable domain expertise, and it has since been released as an open standard, developed in the open on GitHub and documented at agentskills.io, and adopted by roughly three dozen agent products.

Why a folder could become a standard

The reason a folder could turn into a standard is how skills load. Anthropic calls it progressive disclosure, and it has three tiers. The metadata, the name and description, is a few dozen to a hundred tokens per skill and is always in context. The full SKILL.md body, recommended to stay under five thousand tokens, enters context only when a task matches the description. Bundled scripts and references load only when the skill actually references them. A catalog of hundreds of skills costs almost no context until one of them fires. That is what made the format cheap enough to adopt.

The spec is specific about the small things. A skill name uses lowercase letters, numbers, and hyphens, at most 64 characters, and it has to match the folder it lives in. Optional frontmatter fields cover license, compatibility, metadata, and an experimental allowed-tools list. None of this is glamorous. The specificity is the point. A format can only be portable if everyone agrees on what counts as a valid name.

Where it came from

The origin is not in dispute, even if the timing is fuzzy. The format came from Anthropic and was later released as an open standard. The spec is now developed on GitHub under the agentskills organization, where it carries roughly 25,000 stars. I could not find a dated announcement for when it became an open standard; the primary sources describe the sequence without naming a date.

The adoption list

The adoption is the part worth sitting with. The client showcase maintained by the Agent Skills project lists roughly three dozen supporting products: Claude Code, Claude on claude.ai, Gemini CLI, Cursor, GitHub Copilot, VS Code, OpenAI’s ChatGPT and Codex, OpenCode, OpenHands, Block’s Goose, ByteDance’s TRAE, Spring AI, Databricks Genie Code, Pulumi Neo, Hermes Agent, and OpenClaw. By that directory’s count, Anthropic, Google, Microsoft, and OpenAI each ship at least one product that reads the folder format.

A small convention has already formed on top of the spec. Because the standard does not mandate where skills live, a shared directory called .agents/skills/ has emerged so that a skill installed by one compliant client is visible to the others. It is a de facto standard layered on a de jure one.

The enterprise arrival

The enterprise arrival is what makes this more than a developer-tools story. In January 2026 Spring AI shipped a skills implementation for the Java ecosystem, and the framing was explicit: LLM-agnostic, no vendor lock-in, and seamless support for any existing Claude Code Skills. The exact line is that you define your skills once and use them with OpenAI, Anthropic, Google Gemini, or any other supported model. A Java enterprise stack did not build its own format; it adopted someone else’s.

Vercel publishes an official collection of skills at roughly 31,000 stars, bundling deployment, React and Next.js performance, accessibility, and writing rules as installable skills in the Agent Skills format. Platform vendors ship their own skills in the format, which is a stronger signal than reading it.

The momentum numbers point the same way. Vela Partners ranked Agentic Skills Frameworks and Plugins the second fastest-growing open-source AI infrastructure trend in its 2026 scan, adding about 120,000 stars across 24 repositories in 90 days. Vela itself cautions that stars are noisy, and I will come back to that. A separate GitHub trend digest, agents-radar, framed the same moment as a universal skill and context layer coalescing around the agentskills.io standard, with top projects explicitly designed to plug into Claude Code, Codex, Cursor, and Gemini CLI.

What actually changed

Here is the part I think matters, and it is not the star counts. The agent stack is unbundling. For the last two years the question was which framework wins, LangGraph or CrewAI or AutoGen, and which orchestration model to pick, graph or role or chain. That question is still live, and JetBrains still treats it as one with real stakes. The skill layer answers a different question: where does capability live, and who owns it.

Skills standardize capability and knowledge. Frameworks standardize orchestration. They are different jobs, and confusing them is what keeps producing articles about skills killing frameworks. Skills do not kill frameworks. They change where operational knowledge lives, and that place is now a folder instead of a framework.

The enterprise consequence is concrete. A skill is a folder, which means it can be code-reviewed, diffed, and audited like code. That is a direct answer to the governance gap FifthRow points at when it reports that only about 11 to 14 percent of enterprise agent pilots reached production at scale as of March 2026. The skills format turns institutional knowledge into a version-controlled artifact. A procurement playbook, a security team’s triage procedure, and a compliance check can all live in a repo as a skill instead of inside one vendor’s orchestration layer or inside someone’s head.

There is a flip side, and it connects to what we have already written here about agentic security and zero trust. If skills become the distribution mechanism for capability, then who controls them is a security boundary. Skills bundle executable scripts that run with the agent’s privileges. Spring AI’s own post warns that scripts execute directly on your local machine without sandboxing. The integration guide recommends gating project-level skills on trust so an untrusted repository cannot inject instructions into your agent by shipping a folder. That is a real attack surface, not a theoretical one.

What the standard is not

I want to be careful about what this standard is and is not, because the gap between the two is where the risk lives.

The spec standardizes file layout and frontmatter. It does not standardize execution. What a skill can do once loaded, network access, sandboxing, package installs, all of that varies by client. Anthropic’s own documentation notes that custom skills do not sync across surfaces and run under different constraints on claude.ai, the API, and Claude Code. A portable folder is not a portable runtime.

The adoption list is self-reported. The showcase is maintained by the Agent Skills project itself, which is Anthropic-originated, so it is an advocacy view rather than an independent audit. The list is real, but I have not independently confirmed each entry against vendor documentation.

The momentum numbers are stars, not production usage, and Vela says so itself. Star counts measure attention. The star counts in this post are point-in-time signals read in September 2026, not measures of adoption or quality.

Skills also do not exist alone. They sit next to MCP for tools and context and A2A for agent-to-agent communication, both governed by the Linux Foundation, alongside the AGENTS.md and CLAUDE.md project-context convention, so a reader could equally see a maturing stack or fragmentation.

Bottom line

The through-line fits in a sentence. Anthropic invented a folder format for giving Claude expertise, the format became an open standard, and the project’s own showcase now lists reading products from every major vendor. The part worth remembering is what that decouples: the reasoning model from the tools it calls, and the institutional knowledge from the vendor.

I keep coming back to the version-control point, because for the reader this site is written for it is the one that changes a decision. If your operational knowledge lives as a folder in a repo, you can move it between Claude Code, Cursor, and a Java service without rewriting it. If it lives in a framework, it moves when the framework moves. The skill format does not make agents safe or well governed on its own. It makes the thing you would want to govern into an artifact you can see, diff, and audit. That is a smaller claim than most of the marketing around it, and I think it is the one that holds up.

Sources

  1. Agent Skills Overview — agentskills.io
  2. agentskills/agentskills — Agent Skills specification — GitHub
  3. Agent Skills — Claude Platform Docs — Anthropic
  4. Agent Skills Specification — agentskills.io
  5. vercel-labs/agent-skills — GitHub
  6. Agent Skills — Client Showcase — agentskills.io
  7. Adding skills support to your agent — agentskills.io
  8. Spring AI Agentic Patterns (Part 1): Agent Skills — spring.io blog
  9. Top Agentic Frameworks 2026 — JetBrains PyCharm Blog
  10. 19 Open-Source AI Infrastructure Trends 2026 — Vela Partners
  11. AI Open Source Trends 2026-06-23 (Issue #1764) — agents-radar
  12. AI Agent Orchestration Goes Enterprise (April 2026 Playbook) — FifthRow