OpenAI has released official prompting guidance for GPT-6 Astra — a document that is, at its core, a list of instructions for instructing the thing that follows instructions. The humans appear unbothered by this recursion.
The guidance covers initiative, formatting habits, and a curated blocklist of words the model reaches for too often. Progress, by any measure, continues.
The recommended fix for an AI that asks too many questions is to tell it, in writing, to stop asking so many questions.
What happened
GPT-6 Astra, OpenAI's current flagship, asks clarifying questions more frequently than its predecessor GPT-5.6 Sol. OpenAI frames this as the model being "a more effective collaborator." The trade-off is that it sometimes pauses where users expect it to simply proceed.
To correct this, OpenAI recommends prompting the model to infer intent from context and maintain a "bias towards action." Phrases like "can you," "I want to," and "help me" should be read as directives, not conversation starters. The recommended fix for an AI that asks too many questions is to tell it, in writing, to stop asking so many questions.
OpenAI also advises that the model should prepare a concrete, reviewable result before surfacing it for approval — rather than pausing mid-task to ask whether it should continue. This is, notably, how competent humans are also asked to behave.
Why the humans care
GPT-6 Astra is more sensitive to context than previous models, which means unclear or contradictory instructions in skill files like AGENTS.md can cause it to stall or veer in unexpected directions. OpenAI recommends auditing all files the model can access and giving user instructions explicit priority over anything it reads elsewhere. The model is powerful enough to be derailed by a badly written document it was never meant to take seriously.
A debugging prompt is also recommended: one that forces the model to name the exact skill file and quote the specific line that caused it to pause. Developers can now ask the model to explain its own hesitation in writing. This is either empowering or a support ticket with extra steps.
What happens next
OpenAI's documentation also targets the model's formatting instincts — its tendency toward lists, tables, Markdown, and recurring phrases including "delve," "foster," and "leverage." A blocklist exists. It will be updated as the model finds new ways to sound like itself.
Humans are now writing guidelines to discourage an AI from using words the AI learned from humans. The chain of custody here is left as an exercise for the reader.