OpenAI’s documentation for GPT-6 Astra reads less like a spec sheet and more like a behavioral intervention. The model, by default, asks clarifying questions more often than its predecessor GPT-5.6 Sol. That makes it a stronger collaborator, sure, but also one that stalls exactly when users want momentum. The fix, according to the docs, is a prompt that tells Astra to infer user intent and show a “bias towards action” — treating phrases like “can you…” or “help me…” as direct calls to execute, not invitations for follow-up questions.
The deeper issue is context sensitivity. Astra follows longer instructions better than older models but is also more easily thrown by contradictory guidance in skill files like AGENTS.md. OpenAI recommends auditing every context document the model can access and giving user instructions explicit priority. For debugging, the company suggests a prompt that forces Astra to name the exact skill file and quote the specific instruction that caused it to pause or veer off course. That’s a practical acknowledgment that model behavior in production is shaped as much by the files around it as by the model itself.
Then there’s the writing. Astra tends to default to lists, tables, and Markdown formatting — and reuses the same phrases across sessions. OpenAI’s solution is a blocklist of what it bluntly calls “slop words.” The list is specific: “delve into,” “leverage,” “foster,” “it’s worth noting,” “What’s important is,” and the contrastive “This isn’t about X. It’s about Y” construction are all verboten. Made-up hyphenated compounds like “exact-head checks” and “editorial-row layouts” get called out by name. The model should state what it’s doing rather than listing what it won’t do, and avoid summary closers like “In short” or “The simplest mental model is.”
This is the first time a major lab has codified style guidance this bluntly in model documentation. It reflects a real frustration in the developer community: models trained on internet text pick up the internet’s worst stylistic habits, and those habits compound when models generate text at scale. The instruction to avoid “contrastive phrasing” is particularly telling — it’s aimed squarely at the rhetorical tic that makes AI writing feel formulaic. If you’re building on Astra, the message is clear: spend as much time auditing your prompt context as you do tuning your model parameters. The blocklist won’t write your prose for you, but it will keep the model from writing everyone else’s.
