Key takeaways
- Official Google skills catalog adopting the Agent Skills open standard (agentskills.io) and SKILL.md format
- Measurable impact: Gemini API coding accuracy jumped to 87% with Flash and 96% with Pro after adding the gemini-api-dev skill
- Gemini CLI has three-tier skill discovery — workspace, user, and extension skills with clear precedence rules
- Small catalog (1 skill currently) but backed by Google's Gemini CLI infrastructure and the broader Agent Skills ecosystem
FAQ
What are Google Gemini Skills?
A library of SKILL.md-based capabilities for Gemini API development. Currently ships one skill (gemini-api-dev) that teaches agents best practices for building Gemini-powered apps.
How do Gemini Skills compare to Anthropic Skills?
Both use the SKILL.md format from the Agent Skills open standard. Anthropic's catalog is much larger (73K stars, dozens of skills). Google's is newer and smaller (1.8K stars, 1 skill) but deeply integrated into Gemini CLI with three-tier discovery.
Do Gemini Skills work outside Gemini CLI?
Yes. They follow the Agent Skills open standard, so any tool supporting SKILL.md can use them — Cursor, Claude Code, OpenCode, and others.
What Are Google Gemini Skills?
Google Gemini Skills is an official catalog of SKILL.md-based capabilities for Gemini API development . Built on the Agent Skills open standard , it provides agents with up-to-date knowledge about Gemini APIs, SDKs, and best practices that post-date the model's training cutoff.
The core problem it solves: LLMs have fixed training knowledge, but APIs evolve constantly. Skills bridge that gap with on-demand context .
How It Works
Gemini CLI Integration
Gemini CLI discovers skills from three tiers :
- Workspace skills —
.gemini/skills/or.agents/skills/(committed to version control, shared with team) - User skills —
~/.gemini/skills/or~/.agents/skills/(personal, available across all workspaces) - Extension skills — bundled within installed Gemini CLI extensions
Precedence flows top-down: workspace overrides user overrides extension. The agent autonomously decides when to activate a skill based on the request and skill description — only loading full instructions when relevant (progressive disclosure) .
Installation
Skills install via the standard npx skills CLI or Context7:
npx skills add google-gemini/gemini-skills --skill gemini-api-dev --global
Performance
Google published concrete benchmarks :
| Model | Without Skill | With Skill |
|---|---|---|
| Gemini 3 Flash | — | 87% correct API code |
| Gemini 3 Pro | — | 96% correct API code |
This is one of the few published before/after measurements in the skills ecosystem.
Current Catalog
The catalog is small — currently one skill :
- gemini-api-dev — Best practices for building Gemini-powered apps, including thought circulation patterns and SDK usage
Compare this to Anthropic's catalog with dozens of skills across multiple domains.
Strengths
- Measurable impact — Published accuracy benchmarks (87-96%) give concrete evidence that skills work, not just vibes
- Native CLI integration — Three-tier discovery with precedence rules is more structured than most implementations
- Open standard adoption — Uses agentskills.io format, so skills are portable across Cursor, Claude Code, OpenCode, and others
- Progressive disclosure — Only metadata loads initially; full instructions load on activation
Cautions
- Tiny catalog — One skill is barely a catalog; Anthropic has 50x more content
- Not officially supported — Repository disclaimer says "not an officially supported Google product" and isn't eligible for Google's vulnerability rewards program
- Gemini-centric content — The skill itself teaches Gemini API usage, so it's most useful for Gemini developers specifically
- Early stage — 1.8K stars vs 73K for Anthropic's catalog; community adoption is still nascent
Competitive Positioning
| Google Gemini Skills | Anthropic Skills | OpenAI Skills | |
|---|---|---|---|
| Stars | ~1,800 | ~73,000 | ~9,300 |
| Skills count | 1 | Dozens | Dozens |
| CLI integration | Gemini CLI (native) | Claude Code (native) | Codex CLI (native) |
| Standard | Agent Skills (agentskills.io) | Agent Skills (agentskills.io) | Agent Skills (agentskills.io) |
| Cross-platform | Yes | Yes | Yes |
All three now converge on the same open standard — the differentiation is catalog depth and platform-specific integration.
The Tembo Angle
Google joining the Agent Skills standard validates the emerging consensus: skills are becoming the standard unit of agent capability. For orchestration platforms like Tembo, this means skill-aware routing — choosing which agent handles a task based on available skills — becomes increasingly viable as catalogs grow across providers.
Bottom Line
Recommended for: Gemini API developers who want measurably better code generation from Gemini CLI.
Not recommended for: Anyone looking for a broad skills catalog — Anthropic's is far more comprehensive today.
Outlook: The significance isn't the catalog size but the signal: Google, Anthropic, OpenAI, and Cursor all converging on the same SKILL.md standard. Google's contribution is small but their adoption of agentskills.io over a proprietary format strengthens the ecosystem.