Google Cloud Agent Skills 完整指南：从基础到高级云运维

欢迎阅读有关 Google 云代理技能的权威指南。随着软件工程的发展，AI 编码代理正越来越多地被赋予任务，变得自主并能够执行复杂的多步骤云操作。

本系列的目标是了解 Google 云目前提供的官方技能集，从基础开始，并将其应用于一些用例，例如执行安全审计、部署无服务器应用程序或优化 BigQuery 流水线。

在本系列结束时，您应该能够评估其准备情况，并能够利用其中一些代理技能准确且安全地完成工作。

：https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fa9ohfkqgsf7mqc22qoai.png 由 Gemini Notebook 生成

本系列：- 第 1 部分：什么是 Google 云技能（本博客）_- 第 2 部分：中级技能实战（BigQuery 和 Cloud Run）

代理技能是原始 AI 能力与结构化云专业知识之间的关键桥梁。它们是可移植的指令集，教 AI 编码代理如何安全高效地使用特定技术。

与其让代理盲目猜测 gcloud 命令并破坏生产集群，不如技能提供：

这就是我们的理解和预期。归根结底，结果会有所不同，取决于技能的质量。因此，像云操作这样的重要领域，如果使用代理技能，将需要仔细评估。

官方 google/skills 仓库：https://github.com/google/skills 是这些指令集的开源集合。它基于 agentskills.io 上定义的开放标准：https://agentskills.io/，当前（撰写本文时）拥有近 70+ 种技能，涵盖 8 个不同类别：

：https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F06isut608dtvpxdu0k4f.png

如果你将这些技能中大约70条指令/细节同时加载到AI代理的上下文窗口中，代理会变得混乱、运行缓慢且成本高。为了解决这个问题，在处理技能时，代理将使用渐进披露模型。

关于这个话题已经有足够的论述，因此我们在这里不再过多重复。

本节总结了开始使用 Google 技能所需的基础设置。我们将探讨如何使用这些强大的指令集来设置你的代理环境。

在安装任何东西之前，请确保你的计算机上安装了以下内容：

将技能下载到你的计算机上非常简单。请确保已安装 Node.js（v18+）和你的 AI 编程代理。你可以通过 npx skills add google/skills 命令安装需要的特定技能。以下是安装技能的一些方法。

选项A：只安装本教程所需的内容（推荐） 本教程在各步骤和用例中使用了 9 个特定技能。仅安装这些技能可以保持环境专注，这是最佳实践，你应该安装所需的技能，而不是全部安装。

你可以使用交互式 CLI 来安装技能：

从这 9 个技能开始最适合学习。随着探索，你可以随时添加更多。

例如，当我运行第一个命令时，会出现如下交互显示：

接着它会询问你选择全局范围还是项目范围。建议继续使用当前文件夹，即项目范围。这将创建一个 .agents/skills 文件夹，技能将放入其中。

最后，你应该会看到关于安装的全面报告：

这里 ~/agy2-projects/google-skills-tutorial/ 是我运行本教程的文件夹，你的计算机上可能会有所不同。

选项 B：安装附加技能（稍后） 如果你想获取完整目录并希望自行探索，可以使用交互式安装程序：如果你想获取完整目录并希望自行探索，请使用交互式安装程序。

当你运行此命令时，终端中会出现交互选择器。选择你感兴趣的特定技能进行安装和尝试。

：https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fbj7v3g75or67663xfron.png

选项 C：稍后安装单个技能（基于 npx 命令） 如果你发现稍后需要添加额外技能，可以随时单独添加：

手动安装 如果你希望完全控制，可以克隆存储库或手动将技能文件夹放入你的环境中。根据技能的预期作用范围，技能可以安装在两个位置之一：

技能安装位置 当你的代理启动时，它会在多个位置查找技能。它使用特定层次结构解析它们。如果同名技能存在于多个位置，则最靠近工作区的技能优先。

：https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Frgtyrkjn1ctk82wpji3c.png

验证 安装后，验证一切是否正常是一个好习惯。

你可以使用以下命令列出所有当前安装和激活的技能：

你应该会看到类似以下内容：

或者，你也可以查看你的目录结构，以确保 SKILL.md 文件存在：

如果你使用的是 Antigravity CLI，可以从终端启动 Antigravity CLI（agy）。输入 /skills 命令，你应该会看到工作区中列出的技能。

管理技能 CLI 提供了简单的命令，用于随时间管理你已安装的技能：

一旦技能安装完成，它实际上是什么样子？在此步骤中，我们按照 agentskills.io 标准分解技能的结构。

SKILL.md 文件 每个技能的核心是 SKILL.md 文件。这不仅仅是文档，它是 AI 编程代理可执行的指令代码。当代理读取此文件时，它会准确学习如何执行任务、运行哪些命令以及执行哪些安全检查。

让我们查看 sample-skill/SKILL.md 的完整内容，并分解每个部分的作用。

YAML 前置内容 每个 SKILL.md 文件都以 YAML 前置内容块开头。该元数据对于代理的启动阶段至关重要。

Markdown 主体部分 Markdown 主体部分决定了代理的行为。虽然高度灵活，但 Google Cloud 技能通常遵循一个常见的模式：

工作量评估问题：代理在开始之前应向用户提出的澄清性问题。

验证清单：验证任务是否成功的步骤。

可选辅助文件 技能文件夹不仅限于 SKILL.md。它还可以包含辅助目录：

技能结构和处理 下面的图表是一个很好的方式来可视化代理在启动时如何扫描技能的 YAML 前置内容，并随后处理技能加载。

：https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F4m5bvpyu9fev8aoz65ug.png

技能类别表 Google Cloud 技能总体上分为 8 个不同的类别，截至撰写本文时。

：https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fgv0s41w3uv8zs3zir88e.png

现在您已经安装了技能并了解了它们的结构，是时候看看它们的实际应用了。

人工智能编码代理使用渐进式披露模型来高效管理技能：

这可以防止代理的上下文窗口被当前任务不需要的工具指令所淹没。

：https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F1o9hvlfkp5wtzqe5sl8n.png

让我们追踪触发 google-cloud-recipe-onboarding 技能时会发生什么。

用户：“帮助我在我的新项目中开始使用 Google Cloud。”

确保您已经从包含 .agents/skills 文件夹的文件夹中启动了您喜欢的代理（Antigravity、Antigravity CLI 或其他任意代理）。

我建议你暂时不要运行这些提示，而是观察你可以使用哪些自然语言提示，通过这些 Google Agent 技能来操作 Google 云平台。

1. 触发 google-cloud-recipe-onboarding

提示：“我想从零开始设置我的第一个 Google 云项目。”

说明：该代理将指导你完成项目创建、账单关联以及初始 IAM 设置。

2. 触发 google-cloud-recipe-auth

提示：“我如何将本地环境认证到 Google 云？”

Welcome to the definitive guide on Google Cloud Agent Skills. As the landscape of software engineering evolves, AI coding agents are increasingly being given tasks to becoming autonomous and capable of executing complex, multi-step cloud operations.

The goal of this series is to take a look at the current set of official skills being made available by Google Cloud, right from the basics and applying it to a few use cases like performing security audits, deploy serverless applications, or optimize BigQuery pipelines.

At the end of this series, you should be able to assess their readiness and be able to leverage a few of these Agents Skills to get the job done accurately and securely.

：https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fa9ohfkqgsf7mqc22qoai.png Generated by Gemini Notebook

The Series: - Part 1 : What are Google Cloud Skills (this blog) _- Part 2 : Intermediate Skills in Action (BigQuery and Cloud Run)

Agent Skills act as the essential bridge between raw AI capabilities and structured cloud expertise. They are portable instruction sets that teach AI coding agents exactly how to work with specific technologies safely and efficiently.

Instead of an agent blindly guessing a gcloud command and breaking a production cluster, a skill provides:

That is our understanding and expectations. At the end of the day, results will vary, depending on the quality of the Skills. Hence an important area like Cloud Operations , if using with Agent Skills, will have to be evaluated carefully.

The official google/skills repository：https://github.com/google/skills is an open-source collection of these instruction sets. Built on the open standard defined at agentskills.io：https://agentskills.io/, it currently houses (at the time of writing) nearly 70+ skills spanning 8 distinct categories:

：https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F06isut608dtvpxdu0k4f.png

If you were to load all 70 odd instructions/details that are present in these skills into an AI agent’s context window simultaneously, the agent would become confused, slow, and expensive to run. To solve this, when it comes to Skills, Agents will use a progressive disclosure model .

Enough has been written on this topic, so we will not iterate much on that here.

This section summarizes the foundational setup required to start using Google Skills. We will explore how to get your agent environment set up with these powerful instruction sets.

Before we install anything, ensure you have the following installed on your machine:

Getting the skills onto your machine is straightforward. Ensure you have Node.js (v18+) and your AI coding agent installed. You could install specific skills that you need via the npx skills add google/skills command. But here are some ways to install the skills.

Option A: Install Only What This Tutorial Needs (Recommended) For this tutorial, we use 9 specific skills across the steps and use cases. Installing only these keeps your environment focused and is a best practice, you should install what you need, not everything.

You can use the interactive CLI to install skills:

Starting with just these 9 skills is ideal for learning. You can always add more later as you explore.

For example, when I run the first command , I get the following interactive display:

It then asks you for Global or Project scope. It is advisable to go ahead with the current folder i.e. Project scope. This will create a .agents/skills folder and the skills will be placed in that.

Finally, you should see an overall report on the installation:

Here ~/agy2-projects/google-skills-tutorial/ is my folder in which I am running this tutorial. This could be something different on your machine.

Option B : Install Additional Skills (Later) If you want the full catalogue and prefer to explore on your own, use the interactive installer: If you want the full catalogue and prefer to explore on your own, use the interactive installer:

When you run this command, an interactive selector will appear in your terminal. Choose the specific skills that you are interested to install and try out.

：https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fbj7v3g75or67663xfron.png

Option C : Installing a Single Skill Later (npx command based) If you discover you need an additional skill later, you can add it individually at any time:

Manual Installation If you prefer full control, you can clone the repository or manually place skill folders in your environment. Skills can be installed in two locations depending on their intended scope:

Where Skills get installed When your agent starts up, it looks for skills across several locations. It resolves them using a specific hierarchy. If a skill with the same name exists in multiple places, the one closest to the workspace takes precedence.

：https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Frgtyrkjn1ctk82wpji3c.png

Verification After installation, it is good practice to verify that everything is in order.

You can list all currently installed and active skills using:

You should see something similar to this:

Alternatively, you can just peek into your directory structure to ensure the SKILL.md files are present:

If you are using Antigravity CLI, you can launch Antigravity CLI ( agy ) from the terminal. Give it the /skills command and you should see the skills listed in the workspace.

Managing Skills The CLI provides simple commands for managing your installed skills over time:

Once a skill is installed, what does it actually look like? In this step, we break down the structure of a skill following the agentskills.io standard.

The SKILL.md File The heart of every skill is the SKILL.md file. This isn't just documentation, it is executable instruction code for AI coding agents. When an agent reads this file, it learns exactly how to perform tasks, what commands to run, and what safety checks to enforce.

Let’s look at the full contents of our sample-skill/SKILL.md and break down what each section does.

YAML Frontmatter Every SKILL.md begins with a YAML frontmatter block. This metadata is crucial for the agent's startup phase.

Markdown Body Sections The markdown body dictates the agent’s behavior. While highly flexible, Google Cloud skills typically follow a common pattern:

Workload Assessment Questions : Clarifying questions the agent should ask the user before beginning.

Validation Checklists : Steps to verify the task succeeded.

Optional Supporting Files A skill folder isn’t limited to just SKILL.md. It can contain supporting directories:

Skill Structure and Processing The diagram below is a good way to visualize how the Agent at startup scans the Skills YAML frontmatter and then handles the loading of the skill accordingly.

：https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F4m5bvpyu9fev8aoz65ug.png

Skill Categories Table Google Cloud skills are broadly organized into 8 distinct categories, at the time of writing.

：https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fgv0s41w3uv8zs3zir88e.png

Now that you have skills installed and understand their anatomy, it’s time to see them in action.

AI coding agents use a progressive disclosure model to manage skills efficiently:

This prevents the agent’s context window from being overwhelmed with instructions for tools it doesn’t need for the current task.

：https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F1o9hvlfkp5wtzqe5sl8n.png

Let’s trace what happens when you trigger the google-cloud-recipe-onboarding skill.

User : “Help me get started with Google Cloud in my new project.”

Ensure that you have launched your favourite Agent (Antigravity, Antigravity CLI or any other) from the folder in which the .agents/skills folder is present.

I suggest that you don’t run these prompts at the moment and just observe the kind of natural language prompts that you can use to work with Google Cloud Platform via these Google Agent Skills.

1. Trigger google-cloud-recipe-onboarding

Prompt: “I want to set up my first Google Cloud project from scratch.”

Explanation: The agent will guide you through project creation, billing linking, and initial IAM setup.

2. Trigger google-cloud-recipe-auth

Prompt: “How do I authenticate my local environment to Google Cloud?”