OpenAI’s official GPT-6 Astra prompting guide includes a list of banned words!

OpenAI published their official prompting guide for GPT-6 Astra. They even admitted the model is passive by nature (asks for consent, not actions), and provided a list of banned words (e.g., “delve into”, “leverage”, “it’s worth noting”) along with suggestions on how to fix them. For example, they recommend to add “bias towards action” to your system prompt to encourage the model to be more active.

GPT-6 Astra asks more clarification questions than previous versions and seems to prefer returning lists and tables (not always desirable). It’s suggested to always use “plain language”, “active voice”, and “concrete, reviewable results” before presenting anything to the user.

1What This Changes for the Industry

It’s the first time we see a major research lab embedding a “flawed” section into their official documentation (not blog posts or 3rd party tutorials). It means that from now on, a quality model should come with a user manual explaining how to avoid its shortcomings.

They also suggest keeping your test suites small while making changes to code, and using sub-agents. So OpenAI is training the model to be more agentic (speed is more valuable than safety).