How to use the gcloud CLI Guardrails agent skill

What the gcloud skill makes the agent do

Without the skill, an agent may write gcloud commands from memory, and the source describes that memory as stale and prone to hallucinated flags. With the skill loaded, the agent is told to run gcloud help on the exact leaf command (for example, gcloud help compute instances create) before proposing flags or executing anything. Parent group help, such as gcloud help compute, does not count.

The skill also limits how much data comes back. Any list command must include at least one of --limit, --filter, or --format, and the agent is told to use a single-item list first to learn the JSON key names. Execution rules require one command at a time, no pipes, no command substitution, and no redirection, and the --quiet flag on execution commands so runs do not hang waiting for input.

• Validates each leaf command with gcloud help before use
• Requires --limit, --filter, or --format on every list command
• Runs one command at a time with no pipes, $(...) or redirection
• Asks for explicit user approval before denylisted operations
• Tags terminal commands for usage attribution

Install the gcloud skill

Download the ZIP file from this page. Then choose one of the two locations below. For a personal setup that applies to all your projects, unzip it into ~/.claude/skills/. For a single project, unzip it into that project's .claude/skills/ folder.

If you use the Claude apps instead, upload the skill under Settings → Capabilities → Skills.

• Personal: unzip into ~/.claude/skills/
• Per project: unzip into the project's .claude/skills/ folder
• Claude apps: Settings → Capabilities → Skills, then upload

Install and authenticate the Google Cloud CLI first

The skill does not install gcloud for you. The source says that if the gcloud binary is missing, you should follow the official Google Cloud CLI installation guide for your platform. The skill's own rules also say the agent must not install gcloud or its components on its own. It has to tell you what it plans to install and ask first.

You also need access to a Google Cloud project where your account has suitable IAM permissions. The skill assumes gcloud is already authenticated.

• Google Cloud SDK (gcloud CLI) installed and authenticated
• Access to a Google Cloud project with appropriate IAM permissions
• Optional: Cloud CLI remote MCP server for the run_gcloud_command tool

Trigger the skill with natural requests

You do not need a special command to use the skill. It is meant to load when you ask the agent to plan, generate, construct, propose, describe, or execute any gcloud command, including questions about gcloud syntax or flag formatting. Mention your project and the task in plain words, and include the --project flag if you want it preserved in the proposed commands.

The skill does not apply to writing Google Cloud client library code or raw REST or gRPC API requests. Ask for those separately.

• "Using project my-project, list the Cloud Run services in us-central1, showing only name and URL."
• "Plan how to delete the temp Pub/Sub topic in project test-project. Check the syntax first."
• "What flags does gcloud compute instances create accept?"

Examples of what a good session looks like

When a plan is requested, the skill's template has four steps: validate syntax with gcloud help on the leaf command, verify parameters and check whether --dry-run or --validate-only is supported, run a dry run if it is supported, then propose the command. If the command is on the Prohibited Operations denylist, the agent must stop and ask you for authorization. Other commands can proceed under the execution constraints.

For terminal commands, the agent prefixes gcloud with CLOUDSDK_METRICS_REQUEST_ATTRIBUTION set to agent-skills/1.0 (skill:gcloud), unless a more specific product skill such as cloud-run-basics or bigquery-basics applies.

Safety notes: scripts and network access

The skill's package runs scripts and uses the network, so review what you install. The skill itself is written as guardrails for the agent, and it relies on the agent following those rules. The source states that it is not hard technical enforcement.

If you use the optional remote MCP server, the run_gcloud_command tool sends commands to https://cloudcli.googleapis.com/mcp. Do not rely on this skill alone when the agent holds production credentials and no human reviews its actions.

• Runs scripts: yes
• Uses network: yes
• Guardrails depend on the agent following instructions, not on technical enforcement
• Keep human review in place for production credentials

Tips for getting the most from the skill

Ask the agent to show its gcloud help lookup before it proposes flags. If a list output is large, ask it to narrow the result with --limit, --filter, or --format. Keep the skill's reference files alongside SKILL.md, since the agent reads them for CLI usage, MCP usage, and attribution details.

• Ask for the help lookup and dry-run step to be shown explicitly
• Start with a single-item list to see the field names
• Confirm every denylisted action before approving it