GitHub published a practical guide on agent skills, based on a GitHub Universe 2026 session with Procore's Shishir Tewari. The core idea: instead of re-explaining tasks to agents or relying on a colleague's institutional knowledge, you capture repeatable instructions in a skill that you and your agents can reuse — and that can later become part of a team's pipeline.
A skill is a versioned directory centered on a SKILL.md file. Its frontmatter names the task and describes its purpose, and the body explains when the skill applies and how to perform it. Skills are portable across agents, but they don't grant anything themselves: the environment must still provide the tools, credentials, and permissions. A skill can describe how to query a database without granting access to it.
Key takeaways from the guide:
- Define one job per skill, with a clear input and output — for example, a skill that checks AI-generated changes against the original request.
- Avoid self-verification: their Proof Agent skill uses a separate verifier to examine changes, on the principle that self-verification is not verification. Its SKILL.md includes a When to Verify section describing when to invoke the skill.
- Skills are versioned, so an individual's skill can evolve into shared team tooling.
Summary by angular.love. The original source has the full details.