Google Agent Skills is an open-source project that packages Google Cloud domain knowledge into structured instructions to make AI coding agents smarter, safer, and more accurate. The project originated as a cross-functional ‘swarm’ effort before Google Cloud Next 2026, combining Developer Advocates and Technical Writers. Initial reception was strong, with over 15,000 GitHub stars, leading product teams across Google (including Ads) to contribute skills for their own services.

Scaling without losing quality required strict governance. Every skill follows a standardized repository layout. A key architectural principle is to prefer remote MCP tools over CLI or API calls when possible. Public export strips internal assets, ownership, and evaluation suites. Automated CI/CD checks run on every check-in, including linters for metadata and naming, link checkers to prevent hallucinated URLs, and AI-assisted structural validation.

Continuous evaluations run on submit and weekly. On-submit evaluations require authors to provide prompt suites and scoring rubrics. Weekly jobs compare agent performance with and without each skill, measuring accuracy (task completion) and efficiency (token and time usage). Results are placed in a 2×2 matrix to confirm measurable uplift.

The team treats skills as living products with strict ownership: repo maintainers oversee infrastructure, skill owners maintain their skills long-term and update them on API changes or quality degradation. To support contributors, the team built internal authoring tools and agentic workflows using the ADK that run multi-agent loops for self-critique.

Separately, an internal initiative called DevRel Skills encodes team workflows (content transformation, SEO, reporting) into dedicated skills for daily internal use. The post offers links to the repo and two other guide posts, and promises a deeper dive into authoring tools in future articles.