The Quiet Standardization of AI Agent Skills

Anthropic, OpenAI, Google, GitHub, and a growing list of independent developers have all converged on the same file format for customizing AI agent behavior — without any formal standards body, working group, or coordination.

The format is called SKILL.md. Originally created by Anthropic for Claude Code, it has been independently adopted by Codex CLI, Gemini CLI, GitHub Copilot, Cursor, OpenClaw, and 20+ other agents. The same file works across all of them without modification.

How it happened

In early 2025, Anthropic shipped Claude Code with a simple customization mechanism: drop a markdown file called SKILL.md into a directory, and the agent reads it as instructions for a specific task. YAML frontmatter with a name and description. Markdown body with instructions. That's the whole format.

It was not designed as a standard. It solved a specific problem: users wanted to teach their agent recurring workflows without repeating themselves every session.

Then OpenClaw implemented the same format. Then Codex CLI. Then Cursor. Then Gemini CLI. Each implementation was independent — teams read Anthropic's format, found it sufficient, and adopted it rather than inventing something new.

This is unusual. Developer tool ecosystems almost never converge without a formal spec process or a charismatic evangelist pushing adoption. SKILL.md spread because there's nothing to argue about. It's a markdown file with a few YAML fields.

Why this matters

The format war that didn't happen. Compare with YAML vs TOML vs JSON, Terraform vs Pulumi vs CloudFormation, or GGUF vs ONNX vs SafeTensors. Every adjacent space has competing formats. SKILL.md won by being first and simple enough that building an alternative wasn't worth the effort.

Genuine portability. A skill written for Claude Code works in OpenClaw, Codex CLI, Cursor, and Gemini CLI. The instructions are plain English, so there's no agent-specific API to abstract over. Copy the folder, it works.

An interesting question about agent intelligence. If the difference between a mediocre code review and a thorough one is a 20-line markdown file, how much of what we attribute to model capability is actually just better prompting?

The limitations

Discovery is fuzzy — skills activate based on natural language matching, and the agent sometimes loads the wrong one. There's no versioning or dependency management. Cross-agent features (Claude Code's context forking, Codex CLI's openai.yaml sidecar) don't translate.

And there's no trust layer. GitHub has hundreds of thousands of SKILL.md files with no verification. Researchers have already found skills with hidden prompt injection and data exfiltration patterns.

Where this is going

The ecosystem is at the "npm circa 2012" stage. The format exists and adoption is growing, but the infrastructure (discovery, trust, distribution) is still being built.

Things I'm watching: MCP integration letting agents pull skills on demand instead of reading local files. Skill composition (one skill invoking another). And a formal specification — 20+ agents have adopted the format but there's still no spec beyond Anthropic's docs.

Disclosure: I built Agensi, a curated marketplace for SKILL.md skills with security scanning. So I'm biased on the distribution question. But the convergence itself is what I find genuinely interesting.

---

The SKILL.md format reference is at agensi.io/learn/skill-md-format-reference. A curated skill catalog is at agensi.io/skills.