Google Open-Sources 78 Agent Skills for Its Cloud and AI Products

Handing an agent a request like “check the problem with my GKE cluster” is worrying not because it can’t, but because it kind of can — without fully getting it. The commands look real, but the project or region might be wrong, and permission or billing risks can sneak in. So yesterday Google open-sourced its skills repository, organizing Google’s product operation knowledge, execution order, and safety boundaries into instructions an agent can load on demand, task by task.

78 Skills Covering Real Work

The repository currently exposes 78 verifiable skills, grouped by task in the README.

• Getting started on the cloud: from identity authentication and project initialization to solution architecture — solving where an agent should begin once it enters Google Cloud.
• AI and models: Gemini, Agent Platform, model deployment and registration, tuning, RAG, prompt management, evaluation, and alerting.
• Infrastructure and data: GKE from cluster creation through networking, storage, scaling, reliability, security, and upgrades; plus BigQuery, Bigtable, Cloud SQL, AlloyDB, and data lineage.
• Platform and business tools: Cloud Run, Firebase, gcloud, Google Ads, Mobile Ads, and Analytics. More skills for Flutter, Dart, and Cloud Storage live in associated repos.

Agents Won’t Be Flooded by 78 Documents

Agent Skills load on demand. Normally the agent only sees each skill’s name and description, used to judge which tasks might be relevant; only after a match does it read the full SKILL.md, and scripts and reference materials load only when truly needed. That means the repo can keep adding skills without stuffing all of Google Cloud’s docs into context at once. For the user, installing many does not mean paying the same context cost every time.

The Real Value Is in the Steps You “Can’t Skip”

Take the repo’s gcloud skill: it first requires the agent not to assemble commands from memory. Before actually executing, it must verify the exact subcommand with gcloud help; when querying resources it must narrow the scope and fields first, so an unbounded list doesn’t drown the context. It also requires explicitly specifying the project and region — never silently inheriting the terminal’s current config. High-risk actions like IAM changes, deletions, billing operations, and enabling APIs must get explicit human authorization. This is exactly the difference between skills and ordinary knowledge docs: a doc tells you “what the command is”; a skill also tells the agent “what to check before executing, and when it must stop.”

One Task, Written to Be Actionable

The Cloud Run base skill shows another level of granularity. It first distinguishes Service, Job, and Worker Pool, then lists the APIs, roles, and runtime constraints needed for deployment, and explains how to deploy from a container image, source code, or Dockerfile respectively. It even writes in the troubleshooting direction after a failed startup: permission issues point to IAM, health-check failures start with logs, and native-dependency errors mean adjusting the build. The agent gets not a single answer but a path it can keep walking down.

Install in One Command

The repo provides a unified install entry; after running it you can pick the skills you need instead of installing everything at once. If you mainly work with Google Cloud, start with identity auth, gcloud, and the base skills for your products; for Gemini and Agent Platform, keep adding by model calls, deployment, RAG, or evaluation tasks.

npx skills add google/skills

Who Is This For

This repository fits developer and platform teams already using coding agents and wanting them involved in cloud resource queries, deployment, troubleshooting, and configuration. Especially when a team needs a unified execution order, preserved human confirmation points, and less “everyone writes their own prompt from experience,” the value of skills becomes obvious.

Finally: What Are Skills

To keep newcomers from being lost, a quick note on what skills are. This repo adopts the open Agent Skills format. Each capability lives in a folder whose core is SKILL.md: the front matter uses a name and description to tell the agent when to enable it, and the body gives concrete steps, commands, exceptions, and caveats. When needed, a skill can also carry scripts, reference materials, and templates. It is more like an installable, version-controllable work manual than one very long prompt.

GitHub: https://github.com/google/skills