14 months ago, customizing an AI agent meant a markdown file with some rules.
Today: one open standard, four major tools, 30+ adopters, plugin marketplaces, 44,000+ community skills.
December 2025 → March 2026. Three months. That’s how fast the format converged.

The Gap Nobody Talks About

The format converged. That’s the easy part to measure — SKILL.md is supported across Claude Code, GitHub Copilot, Gemini CLI, and OpenAI Codex.
What didn’t converge is everything that matters for production use: governance depth, per-skill controls, and what “supported” actually means in practice.
The Split

Same SKILL.md format. Very different depth of support:
Per-skill tool restrictions (can you lock a skill to specific bash commands?)
- Claude Code: ✓ Yes
- Copilot: ~ Experimental
- Gemini CLI: ✗ No
- Codex: ~ Global only
Per-skill model override (can a skill run on a different model than the default?)
- Claude Code: ✓ Yes
- Copilot: ✗ No
- Gemini CLI: ✗ No
- Codex: ✗ No
Kill switch (can you disable auto-invocation for sensitive skills?)
- Claude Code: ✓ Yes
- Copilot: ✓ Yes
- Gemini CLI: ✗ No
- Codex: ~ Via plugin
Cross-tool skill reading (does the tool read skills from other tools’ directories?)
- Claude Code: ✗ No
- Copilot: ✓ reads
.claude+.agents - Gemini CLI: ✓ reads
.agents - Codex: ✓ reads
.agents+ imports
The format was the easy win. Governance is where the tools diverge.
The Ecosystem

Every major tool now has a skills marketplace:
- Claude Code — official Anthropic-managed marketplace,
/pluginto discover and install - Copilot — Awesome GitHub Copilot: 175+ agents, 208+ skills, 48+ plugins
- Gemini CLI — community registry searchable from your terminal, Google skills repo
- Codex — plugin marketplace launched March 27 with Slack, Figma, Notion, Sentry, Linear integrations
Cross-tool marketplaces are already aggregating 44,000+ skills with security scanning.
Skills aren’t just instructions anymore. They’re distribution units — things you share, install, and version.
The Insight

The format was the easy part. Winning the format war just cleared the field for the next race.
The next race is governance depth and ecosystem density.
Governance depth: Which tool gives organizations the controls they need to deploy skills safely at scale? Per-skill tool restrictions, model overrides, kill switches, audit logging.
Ecosystem density: Which marketplace is worth building for? The value of a distribution channel is proportional to what’s already on it.
Both of these compound. A tool with deep governance captures enterprise accounts. A dense marketplace attracts more skill authors, which attracts more users, which attracts more enterprise accounts.
The format converged in three months. The ecosystem race will take longer.
The AI Basics #2. Next: What Are Plugins?