Building with AI · Building with the protocol stack
You can describe what an Agent Skill is, how the SKILL.md folder format composes, how to author and publish one, and why versioning, signing and trust are now part of the spec.
A Skill is not a tool. A tool is a function the model can call; a Skill is a packaged capability the model can load. Anthropic shipped the Agent Skills open standard in December 2025 — frontmatter, body, dependencies, signed bundle. Within months VS Code, GitHub, Cursor, Goose, Amp, OpenCode and the Codex CLI adopted the SKILL.md folder format. It is closer to a plugin than a prompt.
The pattern matters because the alternative is what we did for two years: copy-paste prompts. "Use this 1,800-token system prompt for invoice extraction." "Here is our refactor-Python-to-async runbook, paste it at the top." That model does not version, does not sign, does not test, and definitely does not scale across vendors. Skills replace it with a folder structure that any compliant runtime can mount.
The four lessons cover what a Skill is, how to compose one with other Skills and an MCP server, how to author one (frontmatter + body + dependencies), and the governance story — versioning, signing, marketplace and trust. The last lesson is the one builders skip and regret: a Skill is executable knowledge; when you trust one, you are trusting whoever signed it. The standard takes that seriously and you should too.
Chapter contains 4 lessons.