Skip to content
gcloud CLI Guardrails

gcloud CLI Guardrails Skill

Google's official skill for driving the gcloud CLI safely from an agent: validate every command against its own help text, cap the output, and refuse the operations that should never run unattended.

GitHub Docs

Install

Install with the skills CLI

One command for Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, Windsurf and dozens of other agents. Needs Node.js; the skill is fetched straight from google/skills.

$ npx skills add google/skills --skill gcloud

Run it in a terminal. The CLI finds the agents installed on your machine, asks which ones to add the skill to, and whether to install it into the folder you ran it from or for all your projects.

Other ways to install

Copy the whole skill folder — SKILL.md and any scripts or references beside it — into ~/.claude/skills/gcloud/, where Claude Code picks it up in every project. Only git is needed.

tmp=$(mktemp -d) && \
  git clone --depth 1 --branch main https://github.com/google/skills.git "$tmp" && \
  mkdir -p ~/.claude/skills && \
  cp -R "$tmp"/skills/cloud/gcloud ~/.claude/skills/gcloud

No scripts or references next to it? Saving the SKILL.md alone as ~/.claude/skills/gcloud/SKILL.md works too.

Using another agent? Its personal skills folder
Codex
~/.agents/skills/
Cursor
~/.cursor/skills/
GitHub Copilot
~/.copilot/skills/
Gemini CLI
~/.gemini/skills/
Windsurf
~/.codeium/windsurf/skills/
OpenCode
~/.config/opencode/skills/

About

  • Mandatory `gcloud help <leaf command>` validation before any command is proposed or run
  • Blocks `list` commands that lack --limit, --filter or --format, with a schema-discovery workflow
  • Denylist for IAM, delete, billing, organizations, KMS and infra-manager apply operations
  • Requires --dry-run or --validate-only first wherever the command supports it
  • Forces --quiet and explicit --project plus location flags for headless, non-interactive runs
  • Bans pipes, redirection, command substitution and chaining so every step stays reviewable
  • Recommends --async with operation polling for long-running operations

The gcloud CLI has thousands of leaf commands and its flags drift between releases, which makes it precisely the kind of surface a language model will confidently get wrong. Google's official skill starts from that premise and builds three defences on top of it.

1. Never trust remembered syntax

The skill's opening instruction is blunt: all pre-existing knowledge of gcloud commands, flags and positional arguments is stale and prone to hallucination. Before proposing or running anything, the agent must call gcloud help <leaf command> for the exact command — validation is explicitly not transitive from parent groups, so gcloud help compute does not license a guess at gcloud compute instances create. Even a plan has to include the help lookup as a step.

2. Never blow up the context window

gcloud ... list on a real project can return megabytes of JSON. The skill forbids any list without at least one of --limit, --filter or --format, and teaches the schema-discovery move that makes projection possible: run the list once with --limit=1 --format=json, read the key paths, then request only those fields. It also pins down filter syntax — prefer : for pattern matching, never quote the right-hand side, treat the whole flag as one string.

3. Never surprise the human

A denylist covers what an agent must not do on its own: IAM policy, role or binding changes, any gcloud * delete, gcloud billing *, gcloud organizations *, gcloud kms *, and gcloud infra-manager deployments apply. Enabling APIs is also off-limits unprompted, since it provisions resources and starts billing. Where a command offers --dry-run or --validate-only, using it first is mandatory.

Headless execution

Two rules exist purely because agents have no TTY. Every command carries --quiet so a confirmation prompt cannot hang the run forever, and every command carries an explicit --project plus region/zone/location flags so gcloud never stops to ask which one you meant — with discovery commands listed for finding those values first. Shell operators are banned outright: no pipes, no redirection, no $(...) substitution, one command at a time, so a human reviewing the transcript can read exactly what ran.

Scope

Use it for anything gcloud; do not use it for Google Cloud client-library code or raw REST/gRPC calls. Part of Google's official agent-skills catalogue at github.com/google/skills, Apache-2.0.

Try it

Once installed, ask your agent something like:

List the Cloud Run services in my staging project that have more than one revision, and show only name and region.

SKILL.md

The instructions your agent loads, as published by Google Cloud.

SKILL.md
name
gcloud
metadata
category: CloudInfrastructureAndServices
description
Provides safety-critical validation, guardrails, and data reduction for gcloud CLI operations across Google Cloud Platform (GCP) services and infrastructure. Use when planning, generating, constructing, proposing, describing, or executing any gcloud CLI commands - including when answering questions about gcloud syntax, or formatting flags. Don't use when writing Google Cloud client library code or raw REST/gRPC API requests.

gcloud CLI Skill for AI Agents

[!CAUTION]

### MANDATORY PRE-CONDITION: EXPLICIT LEAF-LEVEL SYNTAX VALIDATION

All pre-existing knowledge of gcloud commands, flags, flag values, and
positional argument syntax is stale and prone to hallucination.

NEVER propose command parameters, output flag options, execute commands, OR
outline step-by-step plans for any gcloud task before validating leaf-level
syntax via gcloud help <command> (or including leaf-level help lookup as a
mandatory step in the plan).

Mandatory Action Rules:

1. Direct Execution & Code Generation: ALWAYS invoke `gcloud help
<leaf_command> (e.g. gcloud help compute instances create or gcloud
help sql instances create`) before proposing or executing the final
command syntax.

2. Planning & Strategy Queries: When asked for a plan, strategy, or next
steps to achieve a user goal (e.g., *"What is your plan to accomplish
X..."), the response *MUST explicitly include running `gcloud help
<leaf_command>`** as Step 1 of the plan before proposing flags or
executing commands.

3. Non-Transitive Validation: Parent command group help (e.g. `gcloud
help compute`) is not sufficient for leaf-level syntax validation.
Validation must occur at the specific leaf subcommand level.

4. FORBIDDEN Web Search Fallback: NEVER use search_web, web search, or
external documentation search tools for gcloud CLI syntax. `gcloud help
<leaf_command>` is the EXCLUSIVE authorized authority for command
syntax.

5. User Flag & Project Preservation: When proposing intermediate command
steps, ALWAYS preserve all user-specified flags (including
--project=<project_id>) in the proposed response text.

6. Mandatory Plan Template: When generating a plan, the response MUST
copy this exact 4-step structure:

- Step 1: Syntax Validation via gcloud help <leaf_command>
- Step 2: Parameter Verification (confirming required and optional
flags, and explicitly checking if the --dry-run or --validate-only
flag is supported)
- Step 3: Dry-Run Command Proposal (If --dry-run or
--validate-only is supported, there MUST be a --dry-run or
--validate-only invocation before the next step.)
- Step 4: Command Proposal & Authorization (If the command is on the
"Prohibited Operations" denylist, state that autonomous execution is
forbidden, and the user MUST be explicitly asked for authorization to
proceed. If the command is NOT on the denylist, propose or proceed
with execution, while following ALL "Execution Constraints" below.)

This document provides essential guidelines and best practices for AI agents
interacting with the Google Cloud SDK (gcloud CLI). Following these rules is
critical to avoid hallucinated commands, flags, flag values, and positional
argument syntax, prevent destructive actions, and minimize context window usage.

Getting Started
1. Installation

If the gcloud executable is missing, refer to the official
Google Cloud CLI Installation Guide
to install it on the current platform (Linux, macOS, Windows, etc.).

2. Authorization

Authenticate the CLI with Google Cloud. Choose the flow that matches the running
environment:

  • User Account (Interactive): Run gcloud auth login. Follow the browser

prompts to sign in.

  • User Account (Headless Flow): If operating on a terminal without a web

browser (e.g. containers, remote SSH), append the --no-browser flag:
gcloud auth login --no-browser. Copy the URL, sign in on another machine,
and return the authentication code.

  • Application Default Credentials (ADC): To authenticate code calls from

local applications or SDK libraries, set up ADC via `gcloud auth
application-default login (append --no-browser` for headless
environments).

  • Service Account (Best for Detached/Headless Automation): Authenticate

directly using a JSON key file. Ideal for fully automated, background tasks
and pipelines: `gcloud auth activate-service-account
--key-file=path/to/key.json`. Note that some organizations may restrict
access to JSON key files for security reasons.

  • **Service Account Impersonation (Preferred for Local Pair-Programming

Agents)**: Leverage the human developer's existing user credentials to
assume a service account identity. Best for local development assistants to
avoid insecure private keys on human workstations: `gcloud config set
auth/impersonate_service_account SERVICE_ACCT_EMAIL`

Separation of Privilege (Critical): Both service account approaches ensure the
agent's permissions remain strictly distinct from the human user's wide access
limits (enforcing least privilege), and ensure actions are properly audited
under the agent's focused identity. *(Impersonation requires
roles/iam.serviceAccountTokenCreator)*.

For more detailed strategies and authentication types (such as Workload Identity
Federation), see
Authorizing the gcloud CLI.

Core Principles
1. Explicit Command Validation (Mandatory)
  • Action: ALWAYS call gcloud help <command> for the exact command

that is intended to be run (e.g., gcloud help compute instances create).

  • Verify: Ensure the command, flags, flag values, and positional argument

syntax are valid for that specific leaf command before attempting execution
or presenting plans. Validation is not transitive from parent groups.

2. Data Reduction Strategies (Mandatory)

Minimize the volume of data returned by gcloud to save context window space
and reduce latency. DO NOT execute any list command without including at least
one data reduction flag (--limit, --filter, or --format).

  • Projection: Use --format="json(key1, key2, ...)" to select only the

specific fields needed for the task. To understand the advanced projection
and formatting syntax, refer to gcloud topic projections and `gcloud topic
formats`.

  • Limiting: Use --limit=N to cap the number of resources returned.
  • Filtering: Use --filter to narrow down results server-side. Prioritize

: for pattern matching and never quote the right side of the colon. Treat
the entire filter flag as a singular string without quoting or escaping
characters. To study the filter expression syntax, refer to `gcloud topic
filters`.

  • Schema Discovery: Unconstrained resource lists can quickly exhaust the

context window with redundant data. To prevent this, discover a resource's
schema before executing queries. If unsure of the JSON key path for
projecting fields (--format) or filtering (--filter), run the targeted
resource's list command (if supported) with a single-item limit:

    gcloud <GROUP> <RESOURCE> list --limit=1 --format=json

Examine this single instance's JSON structure to safely identify the correct
schema keys before requesting full or filtered datasets.

3. Execution Constraints
  • Single Commands: Execute a single gcloud command at a time. No command

chaining or sequencing.

  • No Shell Operators: Do not use command substitution ($(...)), pipes

(|), or redirection (>, >>, <). This is to increase command safety
and ensure commands are more easily understandable and reviewable by users.

  • Non-Interactive Execution (`--quiet` / `-q`): Pass the --quiet (or

-q) global flag on all execution commands (e.g., `gcloud pubsub topics
delete temp-topic --quiet --project=test-project`). AI agents run in
headless, non-interactive environments without a TTY or stdin input
handler. Without --quiet, commands that prompt for user confirmation (such
as deleting resources, approving defaults, or selecting unspecified regions)
will pause execution indefinitely waiting for input, causing background task
timeouts. Including --quiet forces non-interactive mode, causing gcloud
to automatically accept safe default choices or fail immediately with an
explicit error if required parameters are missing.

  • No Blind Lists: NEVER execute a list command without --limit,

--filter, or --format.

4. Project and Location Scoping (Critical)

To ensure commands are deterministic, non-interactive, and target the correct
environment, they must explicitly provide project and location scoping.

  • Explicit Project Target: Do not rely on active configuration defaults.

Always append --project=<PROJECT_ID> to all resource-manipulating and
querying commands (unless running pure local config commands). This avoids
accidental execution against the wrong project.

  • Prevent Location Prompts: Many Google Cloud resources are regional or

zonal. If the location flag is omitted (e.g., --region, --zone, or
--location), gcloud will trigger an interactive prompt to select a
zone/region. This violates the No Interactivity rule. Always provide
explicit location flags if the command requires them.

  • Location Discovery: If the correct region, zone, or location for a

service is not known, run discovery commands first (remembering to limit
results if there are many):

  • Compute Engine (VMs, Networks):
  • gcloud compute regions list --project=<PROJECT_ID>
  • gcloud compute zones list --project=<PROJECT_ID>
  • Other Services (Standard API Style): Many GCP services utilize a

unified locations list command:

  • gcloud <GROUP> locations list --project=<PROJECT_ID>
  • Examples: gcloud artifacts locations list, `gcloud kms locations

list, gcloud secrets locations list`.

Safety & Guardrails
[!CAUTION] **Destructive actions (delete, update, remove) MUST be explicitly
authorized by the user.** Never invoke them autonomously unless explicitly
instructed to do so in the context of a safe, pre-approved workflow.
Prohibited Operations (Denylist)

NEVER execute the following commands autonomously. These require explicit
human-in-the-loop authorization:

  • Any IAM policy, role, or binding modification (Security): Risk of

privilege escalation, administrative lockout, service disruption, or
unauthorized data exposure.

  • No Proactive API Enabling: Assume necessary APIs are enabled. To prevent

unexpected resource provisioning or billing charges, do not proactively try
to enable APIs. User approval is required to enable any API.

  • *`gcloud delete`** (Destructive): Irreversible resource destruction

(e.g., project deletion) or data wiping.

  • *`gcloud billing `** (Financial): Risk of service disruption or unbounded

costs.

  • *`gcloud organizations `** (Governance): Org-level changes affect security

posture for all users.

  • *`gcloud kms `** (Encryption): Risk of permanently locking data.
  • `gcloud infra-manager deployments apply` (Destructive): Autonomous IaC

execution can destroy managed resources.

Execution Guidelines
  • Dry Run (Mandatory): If the --dry-run or --validate-only flag (or

equivalent) is listed in the command help output, ALWAYS include the flag in
the proposed command or initial execution step. ALWAYS preview changes with
--dry-run or --validate-only prior to actual execution.

  • Long Running Operations: For commands that support it, the --async

flag is highly recommended for long-running operations to avoid blocking the
agentic flow. Note that not every command has an --async flag. For
commands that return an operation ID (whether via --async or by default),
operation status must be polled for completion, if needed for the next step.

  • Non-Interactive Flag (`--quiet`): Include --quiet (or -q) on all

proposed or executed commands to guarantee non-interactive execution without
waiting for TTY confirmation prompts.

Structured Workflows
Discovery Workflow

When asked to perform a task on a service that is unfamiliar:

  1. Invoke Help: Call gcloud help <COMMAND> on the target leaf command

prior to execution.

  1. Traverse Command Tree: Run help on command groups (e.g., `gcloud help

compute or gcloud help`) to discover available subgroups and commands if
the exact command is unknown.

  1. Discover Schema: Run `gcloud <GROUP> <RESOURCE> list --limit=1

--format=json` to inspect JSON keys before constructing filters or
projections. DO NOT execute unconstrained list commands without scoping
flags (e.g., --limit=1) to prevent context window exhaustion.

  1. Enforce Data Reduction: Include data reduction flags (--limit,

--filter, --format) on all command executions.

Quick Reference / Cheat Sheet
TaskCommand Template
Discover Schemagcloud <GROUP> <RESOURCE> list --limit=1 --format=json
Filtered Listgcloud <GROUP> <RESOURCE> list --filter="status:RUNNING"
Specific Columnsgcloud <GROUP> <RESOURCE> list --format="json(name, id)"
Learn Filtersgcloud topic filters
Learn Formatsgcloud topic formats
Learn Projectionsgcloud topic projections
Asynchronous Opgcloud <COMMAND> --async
Check Operationgcloud operations describe <OPERATION_ID>
Common commandsgcloud cheat-sheet
List Regions (GCE)gcloud compute regions list --project=<PROJECT_ID>
List Zones (GCE)gcloud compute zones list --project=<PROJECT_ID>
List Locationsgcloud <GROUP> locations list --project=<PROJECT_ID>

Refer to the
gcloud CLI Scripting Guide
for guidance on using the gcloud CLI in automation.

Install

Covered in the Weekly

Featured

Railway's official agent skill: create projects, provision databases and buckets, deploy, manage variables and domains, and read build failures back — from the CLI, API or MCP server.

Browse all skills →