GPT-6 Astra: OpenAI Details Prompts and Words to Avoid

OpenAI releases guidelines to frame GPT-6 Astra in concrete use cases. The delegation to sub-agents, code testing procedures, a list of formulations to avoid, debugging prompts, and a guided migration command to the model are addressed.

OpenAI frames multi-agent delegation and resumes testing

The model has the capability to assign tasks to sub-agents operating in parallel, but it does so less frequently than expected. OpenAI suggests that developers specify when and how much delegation should occur. Message exchanges between agents may exhibit grammatical errors or spacing issues. Regarding programming tasks, GPT-6 Astra performs thorough checks before completion. In the case of minor changes, this can lead to excessive testing relative to the scale of the change. OpenAI recommends instructing the model to only rerun tests in the event of additional failures or persistent issues. Developers considering a transition to GPT-6 Astra have the option to use Codex with the OpenAI Docs skill to automatically apply suggested changes. The command “$openai-docs migrate this project to GPT-6 Astra” should be used. More detailed versions of the recommended prompts are available on the GPT-6 Astra documentation page.

OpenAI recommends clarifying requests to obtain text

GPT-6 Astra tends to structure its responses in the form of lists, tables, and Markdown, and reuses formulations from one session to another. OpenAI recommends explicitly requesting concise paragraphs, in simple language and active voice, to obtain text. By default, it is advised to write in clear paragraphs, each developing a single main idea. Lists should only be used if the information is parallel, sequential, or more easily comparable, and nested lists should be avoided unless the hierarchy cannot be expressed in prose. The expected style favors simple and direct language, concrete examples, and precise verbs, promoting the active voice. The main point should be explicit from the beginning and then developed with the necessary explanations. OpenAI provides a blacklist of “inappropriate words,” including invented compounds with hyphens like “exact-checks” or “editorial-arrangements,” and recommends indicating what is being done rather than what will not be done. Phrases such as “Conclusion:”, “dive into,” “promote,” “use/leverage,” “it should be noted,” “what’s important is,” “Question? Answer,” “It’s not about X. It’s about Y,” “really,” as well as hyphenated compound adjectives are particularly discouraged. Summarized closures like “In summary: …” or “The simplest mental model is: …” should be avoided in favor of a directly indicated action. OpenAI advises against mentioning what will not be done, what remains unchanged, or how to separate or categorize results, and to ban contrasting formulations like “X, not Y” or “X—not Y.” For technical writing, jargon should be limited to cases where it is genuinely useful, and technical details should only be mentioned if they help illustrate an idea or the work for the user.

Prompts designed to encourage initiative and execution

OpenAI advocates for asking the model to deduce the user's intent from the context and adopt an action-oriented stance. Expressions such as “can you…,” “I want…,” or “help me…” should be understood as requests for action rather than solicitations for follow-up questions. The model is expected to infer the goal and scope of the task from the instructions and the context of the exchange, then act and carry the task to completion. If the user initiates new work or seeks to resolve an existing problem, the model should persistently pursue the desired goal. It should work autonomously toward the user's objective, for example, by creating isolated work trees, resolving merge conflicts, performing read-only actions, or creating draft PRs, unless these actions are clearly destructive or irreversible. The request for approval should be formulated after the model has prepared a concrete and verifiable result, with OpenAI specifying that “the user must approve a concrete and examinable result.” Unsolicited warnings, disclaimers, or safety checklists based on hypothetical risks should be excluded from the prompts.

Context, skill files, and debugging procedure

GPT-6 Astra interprets long instructions better than previous versions while being more responsive to context. Vague or contradictory instructions in skill files, such as AGENTS.md, can cause the model to interrupt work or deviate unexpectedly. OpenAI advises checking all skill files and contextual documents accessible to the model and explicitly highlighting the user's instructions. A debugging prompt is recommended to force the model to identify the precise skill file and reproduce the instruction responsible for the halt or change in direction. If a skill leads to a permission request, an interruption, unfinished work, or a divergence from the user's goal, it is advisable to mention and link the consulted SKILL.md file, cite the relevant instruction, and briefly explain its implementation. OpenAI also recommends distinguishing explicit requirements of a skill from the interpretation of directives.

A framing motivated by observed model behaviors

OpenAI publishes indications on areas where GPT-6 Astra exhibits undesirable behaviors and how to circumvent them. Compared to GPT-5.6 Sol, GPT-6 Astra asks more clarification questions instead of making assumptions. OpenAI presents this feature as a factor for more effective collaboration. However, this can lead the model to stop where users expect it to continue working automatically.