Article Explains GPT-6 Astra Features and Prompting Practices

Summary

The article reviews what it describes as GPT-6 Astra’s main capabilities and offers prompts for controlling its behavior during long-running work. It says Astra can continue other work while an asynchronous tool call is executing, accept a response.steer event during generation, and change reasoning effort within a conversation while preserving earlier cache results. It also describes asynchronous misalignment monitoring for cases such as deleting too much data or exposing unauthorized information, while noting that detection does not undo actions already completed. The article says Astra cannot use a none reasoning setting and that EU data-residency API projects cannot use /fast mode. Its prompting guidance focuses on encouraging initiative while requiring confirmation only before destructive or irreversible actions, clarifying that user instructions take priority over Skills and AGENTS.md when they conflict, and explicitly requesting a natural writing style with fewer formulaic AI expressions. It also recommends specifying when subagents should be delegated and keeping testing proportional to the size of a coding change. The author concludes that these controls may be more useful than vague requests to be “more proactive” or to “remove the AI feel.”