Closing the knowledge gap with agent skills

This post explores how Google DeepMind built an agent skill to close the knowledge gap LLMs face due to fixed training cutoffs.

• •LLMs lack awareness of post-training changes like new SDK versions, model updates, and evolving best practices
• •The gemini-api-dev skill explains the Gemini API feature set, current models, SDKs, sample code, and documentation entry points
• •An evaluation harness of 117 prompts tested skill performance across agentic coding, chatbots, document processing, and streaming tasks
• •Gemini 3 series models saw dramatic improvement with the skill (baseline 6.8% → near 100%), while older 2.5 series benefited less
• •Limitations include lack of automatic skill updates and the potential for stale skill data to cause harm over time