Google Official Agent Skills

What It Does

Google's official agent skills repository provides 13 SKILL.md packs covering Google Cloud products: Gemini API, BigQuery, GKE, Cloud Run, AlloyDB, Cloud SQL, Firebase, Cloud Networking Observability, Cloud Onboarding, Cloud Authentication, and three Well-Architected Framework pillars (Security, Reliability, Cost Optimization). Each skill contains a SKILL.md file with YAML frontmatter and a references/ directory with detailed reference docs. The total pack is 804 KB, 84 markdown files. The skill installs via npx skills add google/skills and is compatible with Claude Code, Cursor, Codex CLI, and any platform that reads the SKILL.md convention.

The Good

Deep reference material, especially GKE. The GKE skill ships with 21 reference files totaling 2,940 lines covering cluster creation, networking, security, observability, scaling, cost optimization, batch/HPC, multi-tenancy, upgrades, storage, reliability, compute classes, inference, app onboarding, backup/DR, IaC, MCP usage, and CLI reference. This is not a surface-level overview. The gke-golden-path.md reference alone defines production-grade defaults (private nodes, Workload Identity, advanced datapath) with explicit trade-off guidance for when you deviate.

Multi-language SDK coverage in the Gemini API skill. The Gemini API skill provides working quick-start code for Python, JavaScript/TypeScript, Go, Java, and C#/.NET. It explicitly warns against deprecated legacy SDKs (google-cloud-aiplatform, google-generativeai) and points to the unified google-genai SDK. That kind of "your knowledge is outdated" callout is exactly what agents need.

Consistent SKILL.md structure across all 13 skills. Every SKILL.md has proper YAML frontmatter with name: and description: fields. Internal markdown links (e.g., Core Concepts) all resolve correctly. The skills follow a predictable pattern: frontmatter, overview, setup steps, reference directory listing, and cross-links to related skills.

Clarifying questions pattern. Skills like google-cloud-recipe-auth and google-cloud-recipe-onboarding include explicit "Clarifying Questions" sections that teach the agent to ask the right questions before jumping to answers. The auth skill asks who is authenticating, where code is running, what the target is, and whether a client library is in use. This is good agent design.

The Bad

No executable scripts or tooling. This is a doc-only skill pack. There are no scripts, no CLI tools, no MCP server binaries, nothing to install beyond markdown files. The skills teach agents *about* Google Cloud, but the agent still needs gcloud, bq, kubectl, and Terraform installed separately. Compare this to googleworkspace/cli (26K stars) which ships an actual CLI tool alongside its skills.

Every skill requires Google Cloud credentials. Not a single skill can demonstrate value without an active GCP project with billing enabled. The onboarding recipe skill mentions a $300 free trial credit, but even that requires a credit card. This creates a high-friction first-run experience for anyone who is not already a GCP user.

Inconsistent directory naming. The google-cloud-recipe-onboarding skill uses reference/ (singular) for its reference directory, while every other skill uses references/ (plural). A minor inconsistency, but the kind of thing that breaks scripts that expect a uniform structure.

WAF skills are question lists, not guidance. The three Well-Architected Framework skills (Security, Reliability, Cost Optimization) are long lists of assessment questions rather than actionable guidance. The Security skill alone contains 70+ questions organized by principle, but no decision trees, no code examples, no specific remediation steps. An agent loaded with this skill would spend more time asking questions than solving problems.

Closed contributions. The CONTRIBUTING.md explicitly states "we are not accepting external pull requests or code contributions." This means errors or outdated patterns can only be fixed by Google engineers internally. Given the pace of GCP API changes, this is a risk.

Smoke Test Results

We ran structural validation against the full cloned repository (804 KB, 84 markdown files). The test checks SKILL.md frontmatter, reference file integrity, internal link resolution, and directory conventions.

Structural validation

$ test -f /tmp/google-skills-review/README.md
✅ README.md exists

$ test -f /tmp/google-skills-review/LICENSE
✅ LICENSE exists

$ test -f /tmp/google-skills-review/CONTRIBUTING.md
✅ CONTRIBUTING.md exists

$ test -d /tmp/google-skills-review/skills/cloud
✅ skills/cloud/ directory

$ find skills -name "SKILL.md" | wc -l
✅ 13 SKILL.md files found (got 13)

$ verify frontmatter on all SKILL.md
✅ All SKILL.md files have YAML frontmatter with name:

$ verify description field on all SKILL.md
✅ All SKILL.md files have description field

$ test -f skills/cloud/google-cloud-recipe-onboarding/references/
❌ Skill dirs have populated references/
 (google-cloud-recipe-onboarding uses singular "reference/" instead of "references/")

$ test -f skills/cloud/gemini-api/SKILL.md
✅ gemini-api SKILL.md exists

$ test -f skills/cloud/bigquery-basics/SKILL.md
✅ bigquery-basics SKILL.md exists

$ test -f skills/cloud/gke-basics/SKILL.md
✅ gke-basics SKILL.md exists

$ test -f skills/cloud/cloud-run-basics/SKILL.md
✅ cloud-run-basics SKILL.md exists

$ find skills/cloud/gke-basics/references -name "*.md" | wc -l
✅ GKE has 21+ reference files (got 21)

$ wc -l skills/cloud/gemini-api/SKILL.md
✅ Gemini API SKILL.md is 100+ lines (got 230)

$ verify all internal [./references/xxx.md] links resolve
✅ No broken internal markdown links

$ find . -name "*.md" -not -path "./.git/*" | wc -l
✅ 84 total .md files (got 84)

$ du -sk .
✅ Repo under 2MB (got 804KB)

Pass rate: 20 of 21. The single failure is google-cloud-recipe-onboarding using reference/ (singular) instead of references/ (plural).

Full structural log →

What the runs tell you

The repo is structurally sound. All 13 skills have correct frontmatter, all internal links resolve, and the reference material is substantial. The one directory naming inconsistency is a minor issue that would only affect scripts expecting a uniform references/ directory name. The real limitation is that this is purely documentation: there is nothing to execute, so the structural validation is the ceiling of what automated testing can verify here.

Setup Walkthrough

1. Install the skills via the npx skills CLI:

 npx skills add google/skills

This presents an interactive menu where you select which of the 13 skills to install.

1. Alternatively, clone the repo directly for inspection:

 git clone --depth 1 https://github.com/google/skills.git

The shallow clone is 804 KB.

1. No post-install configuration needed for reading the skills. For using them with Google Cloud, you need gcloud CLI installed and authenticated, a GCP project with billing enabled, and the relevant APIs enabled (e.g., gcloud services enable bigquery.googleapis.com).

Alternatives

1. googleworkspace/cli (26K stars) - Google's other official skills repo. Ships an actual CLI tool for Drive, Gmail, Calendar, Sheets, and Docs alongside agent skills. Covers Google Workspace rather than GCP infrastructure. Prefer this if you need executable tooling.
2. anthropics/skills (135K stars) - Anthropic's official skills. Broader scope, not Google-specific. Includes executable scripts alongside documentation. Better as a general-purpose agent skill pack.
3. obra/superpowers (193K stars) - The dominant agent skills framework. Platform-agnostic methodology with structured workflows. Not Google Cloud specific, but covers software development practices that complement GCP-specific knowledge.