GitHub - google-gemini/gemini-skills: Skills for the Gemini API, SDK and model/agent interactions
Key Points
- 1This repository introduces "skills," a lightweight technique to provide up-to-date context and best practices to large language models, addressing their inherent knowledge gaps due to fixed training data.
- 2Evaluations show these skills significantly improve an agent's ability to generate correct Gemini API code, achieving 87% with Gemini 3 Flash and 96% with Gemini 3 Pro.
- 3The project offers various specialized skills, including `gemini-api-dev` and `vertex-ai-api-dev`, which can be easily installed and browsed using Vercel or Context7 command-line interfaces.
The paper introduces the "Gemini API skills" GitHub repository, which serves as a curated library of "skills" designed to enhance the capabilities of agents interacting with the Gemini API, SDK, and underlying models. The fundamental problem addressed is the inherent knowledge gap of large language models (LLMs), which, due to their static training data, lack up-to-date information regarding rapid software development changes, new libraries, evolving best practices (e.g., thought circulation), and API/SDK modifications.
The core methodology revolves around the concept of "skills" as a lightweight technique for injecting relevant, dynamic, and specialized context into LLM-powered agents. These skills bridge the aforementioned knowledge gap by providing agents with the most current and accurate information pertinent to specific development tasks. While the specific technical mechanism for context injection (e.g., direct prompt augmentation, Retrieval-Augmented Generation (RAG), or guiding tool/function calling) is not explicitly detailed, the objective is to significantly improve the LLM's ability to generate correct and best-practice-aligned code and responses.
Empirical evaluations cited in the paper demonstrate substantial performance improvements: the integration of these skills elevated an agent's capability to generate accurate API code following best practices to 87% when utilizing Gemini 3 Flash and a notable 96% with Gemini 3 Pro.
The repository includes several specialized skills:
gemini-api-dev: Focuses on best practices for general application development using the Gemini API.vertex-ai-api-dev: Tailored for building Gemini-powered applications on Google Cloud Vertex AI, encompassing the Gen AI SDK, tool utilization, multimodal generation, caching strategies, and batch prediction techniques.gemini-live-api-dev: Designed for real-time, bidirectional streaming applications, covering WebSocket-based audio/video/text streaming, voice activity detection, native audio features, function calling, and session management.gemini-interactions-api: Provides context for various Gemini interaction modalities, including text generation, multi-turn chat, streaming, function calling, structured output generation, image generation, Deep Research agents, handling deprecated model guardrails, and supporting both Python and TypeScript SDKs.
Installation of these skills is facilitated via two command-line interfaces: the Vercel skills CLI (npx skills add google-gemini/gemini-skills) and the Context7 skills CLI (npx ctx7 skills install /google-gemini/gemini-skills).
The project is explicitly stated as not being an officially supported Google product and is ineligible for the Google Open Source Software Vulnerability Rewards Program.