OpenAI’s developer guidance for GPT-6 Astra focuses on several predictable failure modes: asking too many clarifying questions, stopping before a task is finished, following conflicting context and producing repetitive, heavily formatted prose. The company supplies prompt patterns intended to make those behaviors easier to control.
For autonomous work, OpenAI recommends explicitly telling Astra to infer intent from the conversation and continue until it produces a reviewable result. The model should prepare that result before requesting approval for a consequential next step. Developers who want parallel work must also state when and how much to delegate, because the model may otherwise underuse subagents.
Longer instructions work better than on earlier models, according to the guidance, but Astra is also sensitive to contradictions in files such as AGENTS.md. OpenAI suggests auditing every context file and using a debugging instruction that requires the model to identify the exact file and rule that caused it to pause or change direction.
The writing recommendations favor plain paragraphs, active voice and concrete language over stock transitions, unnecessary headings and recurring phrases. For coding tasks, teams may also need to limit repeated test runs when a small change does not justify a large suite. These prompts shape behavior rather than change the model’s underlying capabilities, so developers still need task-specific evaluation.