GPT-6 Astra, OpenAI's newest model (we covered what it is and does here), behaves differently enough from earlier models that some prompting habits that worked before stop working as well. OpenAI published its own guidance on this, but it's written for developers calling the API (the way a program talks directly to a model, instead of through the chat window you type into), full of terms like reasoning effort and prompt cache options that mean nothing if you're just using ChatGPT.com or the app. Here's what changed, translated into things you can copy and paste today.
Most of what changed comes down to how directly you need to state what you want. Astra checks in with you more than earlier models did, and it defaults to bullet points and tables even when you'd rather read plain sentences. None of the fixes below require touching a line of code. If you do write scripts or build automations, there's a short section on that too.
The quick answer
- Astra checks in with you more before acting. If you want it to make a call and go, say so directly.
- It defaults to bullet points and tables. Ask for plain paragraphs when that's what you want.
- If you've saved custom instructions or built a Project, conflicting guidance can make it hesitate. Tell it which instruction wins.
- If you use it for code or automations, it tends to write more tests than the change needs, and a few API settings changed names.
Why GPT-6 Astra keeps asking clarifying questions
Astra asks clarifying questions more often than earlier models. Where GPT-5.6 Sol might have picked a reasonable default and run with it, Astra is more likely to stop and ask which option you want first. That's useful when the stakes are high, like sending an email on your behalf, and slows you down when they're not.
If you want it to stop checking in, say so in the moment. OpenAI's own guidance to developers makes a related point: when you ask "can you help me with X," Astra is meant to treat that as permission to start, not a yes-or-no question about whether it's capable. If you notice it responding with something like "yes, I can help with that, want me to?" instead of just doing the thing, telling it plainly to skip the check-in fixes it.
How to stop it from defaulting to bullet points and tables
Astra's default writing style leans on lists and tables more than earlier models did. That's fine when you're comparing options, but it gets in the way when you want an explanation written out in sentences: an email, a caption, a story. Ask for it directly, and Astra follows a direct style instruction more precisely than earlier models did, so this works better now than it used to.
What to do if it ignores your custom instructions
If you use ChatGPT's custom instructions (the settings where you tell it how to behave across every conversation) or a saved Project with its own instructions, Astra follows those more precisely than earlier models. That's mostly good. It can also mean Astra hesitates or gives you a strange answer when what you're asking right now conflicts with something you saved weeks ago and forgot about. When that happens, say it plainly.
Using GPT-6 Astra for code and automations
If you use ChatGPT, or the API behind it, to write scripts, fix code, or build workflows in a tool like n8n, two things changed.
First, Astra tends to write more tests than a small change needs. If you only changed one function, say so.
Second, if you or a tool you use calls the API directly rather than going through the ChatGPT app, a few settings changed names. temperature and top_p, older settings that controlled how predictable or varied a response was, aren't supported on Astra. In their place is a setting called reasoning effort (how much time the model spends thinking before it answers, traded off against speed), set to low, medium, or high. If you're maintaining an n8n workflow or a script that calls the OpenAI API and it stops working after switching to Astra, that setting name is the first thing to check.
What to change in your prompts today
None of this requires relearning how to write a prompt. The one habit worth carrying into every Astra conversation: say exactly what you want instead of leaving it implied. Tell it to act instead of ask. Tell it plain paragraphs instead of bullets. Tell it which instruction wins when two conflict. Astra follows direct instructions more precisely than earlier models did, and vague ones are exactly where the extra check-ins and default formatting show up.
For everything else that changed in this release, we broke down GPT-6 Astra's full capabilities separately. And if you want a reusable framework for prompting any model, not just this one, ITK's own Universal Prompt Engineering Template is built for that.
