GitHub is recommending “agent skills” as a way to preserve repeatable engineering procedures outside individual chats. A skill is a versioned directory centered on a SKILL.md file that defines a bounded job, when it applies, and the evidence or output it should produce.
The format separates reusable judgment from project-specific state. GitHub’s examples keep database names, schemas and platform choices in a reviewable manifest rather than hardcoding them into instructions. Several skills can then share that manifest while handling distinct steps such as discovery, building and linting. Updating one responsibility does not require rewriting the whole workflow.
Portability has a firm boundary. A skill can explain how to query a database, but the agent host must still provide the database tool, credentials and permission. Stable interfaces, long-running state, external policy enforcement and detailed audit trails may be better handled by an application rather than a skill.
GitHub also recommends independent verification for AI-generated work. Its example verifier checks changes against the original request and supplies evidence instead of asking the same agent to grade its own answer. Teams adopting the pattern still need human review where discovered repository habits could be mistaken for approved business rules.