Оксана СтепураAI Eng
8 September 2026, 09:47
2026-09-08
GPT-6 Astra asks questions more often and listens better to instructions. Ukrainian AI expert gives advice on how to prompt her
AI Solutions Architect Maksym Valin paid attention to OpenAI's recommendations for prompting the new GPT-6 Astra and collected some tips on how to better set tasks for it. In particular, he noted that the model is more likely to ask the user if it lacks important information. Also, old or contradictory instructions in the context can have a stronger impact on its behavior.
AI Solutions Architect Maksym Valin paid attention to OpenAI's recommendations for prompting the new GPT-6 Astra and collected some tips on how to better set tasks for it. In particular, he noted that the model is more likely to ask the user if it lacks important information. Also, old or contradictory instructions in the context can have a stronger impact on its behavior.
He talked about this on LinkedIn,based on a guide from OpenAI.
"Nobody banned the old 'You are an expert with 20 years of experience...'. It's just that now a clear description of what exactly needs to be done and how the model should work is much more useful," Valin noted.
Astra has become more cautious where data is lacking. If there is no information that could significantly affect the result, the model will more often ask a clarifying question instead of thinking something up on its own. So Valin advises to immediately explain to the model how much autonomy it can act. For example, if there is no essential information, make an educated guess and move on, and ask again only when the user's answer can really change the result.
The developer also advises immediately describing the final result. Instead of “analyze the document,” Valin suggests, for example, asking the model to find 5 key conclusions, risks, contradictions, and issues that need additional verification. For large tasks, you can also determine when the work is considered complete. For example, after checking all sources, drawing out contradictions separately, and preparing the result in a form that can be used.
Astra also follows long instructions better, but because of this, it can react more strongly to rules that are already in its context. So it is worth checking skills, AGENTS.md and other available model files for old or conflicting instructions. If two rules conflict, the model may stop working or go in the wrong direction the user expected.
And Astra, by default, tends to gravitate toward detailed answers, lists, tables, and formatting, so the company recommends being specific about the structure you want. So instead of telling the model to “write naturally,” it’s better to explain what “natural” means. For example, ask it to write in coherent paragraphs, use simple words, start with the main idea, and only use lists when they really help you compare information.
In addition, OpenAI explicitly recommends banning AI clichés. In an example prompt, the company lists the following undesirable expressions: delve, foster, leverage, it's worth noting, importantly, as well as the construction "This isn't about X. It's about Y."
Additionally, when working with code, Astra tends to test the result thoroughly before considering the task complete. For large changes, this is useful, but for a small fix, the model may run more checks than necessary. Therefore, developers are advised to separately specify the required level of testing. For example, perform checks that are appropriate for the scope of the change, and do not expand them without new errors or other reasons.
It’s a similar story with subagents. Astra can split tasks between multiple agents working in parallel, but OpenAI notes that the model may do this less often than a particular developer needs. If parallel work is important, this should also be explicitly stated in the instructions.
So Valin suggests the following prompt template for Astra: Task → context → autonomy → readiness criteria → format → style.
“The stronger the models become, the less prompting feels like searching for a secret spell and more like a well-defined technical task,” the developer concluded.
Previously, dev.ua wrote that OpenAI has released a collection of over 300 ready-made prompts for working with ChatGPT.
