# Gemma AI Skills `gemma-skills` is Google's open collection of [[AI Agent Skills]] for working with [[Gemma]] models. Apache 2.0, around 960 stars, and carrying the usual disclaimer: not an officially supported Google product. Two skills so far: - **gemma-dev** — building applications with Gemma, plus general questions about the models: prompt structure, capabilities, model selection, deployment. - **gemma-trainer** — training, fine-tuning and adapting Gemma on local hardware. SFT, DPO, RLHF, reward modeling. ## What's actually inside I looked at `gemma-dev` because a skill is only as good as its contents, and this one is denser than the README suggests. Six sections: 1. **Prioritize app tooling** — point developers at high-level SDKs, not raw ML libraries, unless they explicitly asked for training or research work. Good default. 2. **Model selection guide** — [[Gemma 4]] variants (26B A4B, 31B, 12B, E2B, E4B, thinking mode), legacy Gemma 3 sizes, and the task-specific ones ([[EmbeddingGemma]], [[ShieldGemma 2]]). 3. **Deployment workflows** — use case mapped to stack: [[Gradio]] plus Transformers for prototyping, [[Transformers.js]] in the browser, [[Vertex AI]] in the cloud, [[MLX]] on Apple Silicon. 4. **Multi-token prediction** — speculative decoding with the matching assistant model repos. 5. **Quantization-aware training** — QAT formats for [[llama.cpp]], [[vLLM]], mobile, and other engines, with the official [[HuggingFace|Hugging Face]] collections. 6. **Documentation lookup** — try the [[Model Context Protocol (MCP)|MCP]] server's `search_documentation` tool first, fall back to fetching official docs. It also ships asset files: a Gradio app, a transformers.js app, a Vertex AI app. Working examples, not prose. ## Install Two CLIs, both `npx`: ```bash npx skills add google-gemma/gemma-skills --skill gemma-dev --global npx ctx7 skills install /google-gemma/gemma-skills gemma-dev ``` The first is [[Vercel]]'s skills CLI, the second is [[Context7]]'s. Neither is Google's own. That's the notable bit. ## Why this matters A model vendor publishing skills for its own models is the pattern I expect to become normal. Documentation tells a human how to use your model; a skill tells an agent how, and the agent is increasingly the one reading. Google shipping `gemma-dev` is the same move as any library shipping typings. Two things stand out to me. First, Google published to a plain GitHub repo in the same SKILL.md format everyone else uses, then let third-party CLIs handle installation. No proprietary format, no Google-only runtime. See [[AI Skill Distribution]] and [[AI Skill Portability]] for why that's the right call. Second, the skill encodes decisions, not facts. "Use high-level tooling unless asked otherwise" and "prefer Gemma 4 over Gemma 3" are opinions a vendor is uniquely qualified to hold. That's the part documentation is bad at and skills are good at. Worth reading as a reference implementation if you're writing skills for your own tool or library. ## References - Repository: https://github.com/google-gemma/gemma-skills - Gemma documentation: https://ai.google.dev/gemma/docs - Gemma models on Hugging Face: https://huggingface.co/google - Vercel skills CLI: https://github.com/vercel/skills - Context7 skills CLI: https://context7.com ## Related - [[Gemma]] - [[Gemma 4]] - [[Google]] - [[Google DeepMind]] - [[AI Agent Skills]] - [[AI Skill Best Practices]] - [[AI Skill Distribution]] - [[AI Skill Portability]] - [[Claude Code Skills]] - [[Obsidian AI Skills by Steph Ango]] - [[My Published AI Skills]] - [[Context7]] - [[HuggingFace]] - [[EmbeddingGemma]] - [[ShieldGemma 2]]