Copilot Code Review Adopts Anthropic's SKILL.md Convention

GitHub made agent skills and MCP server support generally available in Copilot code review on July 29.

The rollout reaches every Pro, Pro+, Business, and Enterprise Copilot user, and every MCP call a skill makes during review is capped read-only. A skill built assuming full tool access won't behave the same here.

Skills load from a file at .github/skills/name/SKILL.md. Anthropic's own convention, invented for Claude Code. Now running inside Microsoft's product.

GitHub and Playwright MCP servers are on by default. Comments in review now carry attribution, so you can tell whether a finding came from a skill or a server.

The frontmatter conventions are still underspecified, though. I've hit ambiguity around required fields myself, and this format doesn't resolve that yet.

The same week, two more projects landed on the identical file shape without any coordination. refly-ai/refly, 7,469 stars and pushed July 29, calls itself the first open-source builder for agent skills. It ships a separate registry where skills get versioned and forked instead of staying one-off prompts, and exports to Claude Code, Cursor, and Codex. rohitg00/pro-workflow, 2,710 stars, installs one skill bundle into what it claims are 32-plus agents through a single cross-agent installer.

Three unrelated projects agreeing on Markdown-plus-frontmatter is how a format actually wins. Nobody held a meeting about it.

Syntax versus semantics

My bet: SKILL.md becomes the default export for anything that generates agent instructions, the way OpenAPI did for HTTP APIs. But the behavior underneath doesn't travel with the file. A skill written for Claude Code's full tool access won't act the same read-only inside Copilot's review sandbox. That gap decides whether this becomes a real standard or splinters into per-vendor dialects that happen to share a filename.