OpenAI is expanding the functional range of its models by integrating "Skills" directly into the OpenAI API, streamlining how developers add custom capabilities to their agents. This update allows for a "shell" tool environment where models can execute specific scripts or functions on the fly. Instead of requiring complex external hosting or manual pre-registration of functions, developers can now bundle their custom tools as zip files and transmit them as inline base64-encoded data within a single JSON request.

Simon Willison (tech developer and blogger) demonstrated this capability using a script targeting a "gpt-5.2" model. In his example, a word-counting skill was packaged and sent directly to the model, which then executed the code within an automated container environment. This approach significantly reduces the friction of expanding a model's utility, effectively allowing it to carry its own toolbox for specific tasks without persistent infrastructure.

The workflow also highlights the rising importance of specialized developer tools. Willison utilized Claude Code (Anthropic’s command-line interface for coding) alongside Showboat (a tool for demonstrating agent actions) to explore and document the API’s new features. This interplay between various AI coding agents and demonstration tools suggests a future where API exploration and software integration are increasingly handled by AI-to-AI interactions, leaving humans to oversee high-level architecture and intent.

OpenAI is expanding the functional range of its models by integrating "Skills" directly into the OpenAI API, streamlining how developers add custom capabilities to their agents. This update allows for a "shell" tool environment where models can execute specific scripts or functions on the fly. Instead of requiring complex external hosting or manual pre-registration of functions, developers can now bundle their custom tools as zip files and transmit them as inline base64-encoded data within a single JSON request.

Simon Willison (tech developer and blogger) demonstrated this capability using a script targeting a "gpt-5.2" model. In his example, a word-counting skill was packaged and sent directly to the model, which then executed the code within an automated container environment. This approach significantly reduces the friction of expanding a model's utility, effectively allowing it to carry its own toolbox for specific tasks without persistent infrastructure.

The workflow also highlights the rising importance of specialized developer tools. Willison utilized Claude Code (Anthropic’s command-line interface for coding) alongside Showboat (a tool for demonstrating agent actions) to explore and document the API’s new features. This interplay between various AI coding agents and demonstration tools suggests a future where API exploration and software integration are increasingly handled by AI-to-AI interactions, leaving humans to oversee high-level architecture and intent.