OpenAI 的模型文档详细说明了 GPT-6 Astra 在哪些情况下容易出现不符合预期的行为，以及开发者如何规避这些问题。
据 OpenAI 称，与 GPT-5.6 Sol 相比，GPT-6 Astra 更倾向于主动提出澄清性问题，而不是自行做出假设，这使其成为“更高效的协作者”。但代价是，该模型有时会在用户期望它继续推进的地方停下来。
为了推动模型展现出更多主动性，OpenAI 建议在提示词中告诉模型，要根据上下文推断用户的“意图”，并表现出“行动倾向”。像“你能……”、“我想……”或“帮我……”这样的表述应被视为行动指令，而不是邀请模型提出后续问题。
你应当根据指令和此前的对话上下文推断用户的意图和任务范围。你的任务是表现出行动倾向，并坚持完成用户意图中的任务，直到任务完成。如果用户表达了完成新工作或解决现有问题的意图，请持续努力，直到实现用户的目标。应独立朝着用户的目标工作（例如，创建独立的工作树/检出分支、解决合并冲突、执行只读操作、创建草稿 PR 等），除非这些操作明显具有破坏性或不可逆。
模型应等到已经准备好具体、可审查的结果后，再请求用户批准。OpenAI 通过这样的提示词引导模型：“用户应该是在批准一个具体、可审查的结果。”基于假设性风险的主动警告、免责声明或安全检查清单应从提示词中移除。
GPT-6 Astra 在遵循较长指令方面优于其前代模型，但对上下文也更为敏感。技能文件（如 AGENTS.md）中不清晰或相互矛盾的指令可能导致模型中断工作或意外偏离方向。OpenAI 建议审计模型可访问的所有技能文件和上下文文档，并明确赋予用户指令最高优先级。
OpenAI 还推荐使用一种调试提示词，强制模型指出导致其暂停或改变方向的具体技能文件名，并引用具体指令内容。这有助于开发者将意外行为追溯到其根源。
如果某项技能导致你请求许可或确认、暂停、让所请求的工作未完成，或偏离用户的意图，请指出并链接到你读取的确切 SKILL.md 文件，引用相关指令，并简要说明其适用方式。请区分技能的明确要求与你对准则的解读。
“Delve”、“foster”、“leverage”以及其他“口水词”
GPT-6 Astra 倾向于使用列表、表格和 Markdown 格式来组织回复，并在不同会话中重复使用相同措辞。OpenAI 对塑造模型的写作风格有具体指导。如果你想要散文式叙述，请明确告诉模型使用平实语言和主动语态撰写简洁段落。
默认情况下，使用清晰、简洁的段落，每段围绕一个中心思想展开。仅当信息确实具有并列、递进关系或更易于比较时才使用列表，除非层级关系无法用散文清晰表达，否则避免使用嵌套列表。使用简单直接的语言：熟悉的词汇、具体的例子和准确的动词。优先使用主动语态和直接陈述。尽早阐明要点，然后用读者所需的解释和细节加以展开。让每句话都承接前一句。展开论述重要的观点，并提供足够的证据使其具有实用性。
一份典型 AI 用语黑名单也能有所帮助。OpenAI 称其为“口水词”。像“exact-head checks”或“editorial-row layouts”这类生造的连字符复合词也应避免。模型应说明自己在做什么，而不是罗列自己不会做什么。
避免使用诸如结论中的“Conclusion:”、“delve into”、“promote”、“use/leverage”、“it’s worth noting”、“what’s important is”、“Question? Answer”或“This isn’t about X. It’s about Y”、“really/truly”等口水词或短语，以及复合描述和连字符形容词。不要使用“In short:...”、“The simplest mental model is:...”这类总结性收尾语句。直接陈述预期行动。避免提及你不会做什么、哪些保持不变，或你将如何区分或分类结果。不要使用“X, not Y”或“X—not Y”这类对比性措辞，这会引入用户并未要求的、多余的替代方案。避免使用“exact-head checks”和“editorial-row layouts”等生造的复合词、含糊的限定词和套话式过渡语；使用简单的动词和介词直接表达实际关系。
在技术写作方面，OpenAI 建议只在确实有帮助的情况下使用术语：“优先使用平实的语言而非术语，并且只在有助于向用户阐明某个想法或你的工作时，才引用技术细节。”
子智能体委派不足，测试规模膨胀
该模型可以将工作移交给并行运行的子智能体，但这样做的频率低于预期。OpenAI 表示，开发者应该明确说明它应该在何时以及以多大程度进行委派。智能体之间的消息也可能包含语法或空格错误。
在编码任务中，GPT-6 Astra 在收尾前会运行全面的测试。对于小改动，这可能意味着测试套件的规模与实际工作严重不成比例。OpenAI 建议告诉模型，仅在出现新的失败或未解决的问题证明有必要时，才重新运行测试。
这些提示词的更详细版本可在 GPT-6 Astra 模型文档页面上找到。想要切换到 GPT-6 Astra 的开发者可以使用带有 OpenAI Docs 技能的 Codex 来自动应用建议的更改：$openai-docs migrate this project to GPT-6 Astra
OpenAI's model documentation spells out where GPT-6 Astra tends toward unwanted behavior and how developers can work around it.
GPT-6 Astra asks clarifying questions more often than GPT-5.6 Sol instead of making assumptions on its own, according to OpenAI, making it a "more effective collaborator." The trade-off is that the model sometimes stops where users expect it to keep going.
To push it toward more initiative, OpenAI recommends a prompt telling the model to infer the user's "intent" from context and show a "bias towards action." Phrases like "can you...," "I want to...," or "help me..." should be treated as calls to act, not invitations for follow-up questions.
You should infer the user’s intent and the scope of the task from the instructions and the conversation context so far. Your task is to demonstrate a tendency to act and to follow through on the user’s intended task until completion. If the user expresses the intention to complete new work or resolve an existing problem, continue working persistently until the user’s intended goal is achieved. Work independently toward the user’s goal (e.g., create isolated work trees/checkouts, resolve merge conflicts, perform read-only actions, create draft PRs, etc.), unless the actions are clearly destructive or irreversible.
The model should wait to ask for approval until it has already prepared a concrete, reviewable result. OpenAI prompts it this way: "The user should be approving a concrete, reviewable result." Unsolicited warnings, disclaimers, or safety checklists based on hypothetical risks should be dropped from prompts.
GPT-6 Astra follows longer instructions better than its predecessors but is also more sensitive to context. Unclear or contradictory instructions in skill files like AGENTS.md can cause the model to block work or veer off unexpectedly. OpenAI recommends auditing all skill files and context documents the model can access and giving user instructions explicit priority.
OpenAI also recommends a debugging prompt that forces the model to name the exact skill file and quote the specific instruction that caused it to pause or change direction. This helps developers trace unexpected behavior back to its source.
If a skill causes you to ask for permission or confirmation, pause, leave requested work unfinished, or diverge from the user's intent, name and link to the exact SKILL.md file you read, quote the relevant instruction, and briefly explain how it applies. Distinguish explicit skill requirements from your interpretation of guidelines.
"Delve," "foster," "leverage" and the rest of the slop words
GPT-6 Astra tends to structure responses with lists, tables, and Markdown formatting, and reuses the same phrases across sessions. OpenAI has specific guidance on shaping the model's writing style. If you want prose, tell the model explicitly to write concise paragraphs using plain language and active voice.
By default, use clear, concise paragraphs, each developing a single main idea. Use lists only if the information is truly parallel, sequential, or more easily comparable, and avoid nested lists unless the hierarchy cannot be clearly expressed in prose. Use simple, straightforward language: familiar words, concrete examples, and precise verbs. Favor the active voice and direct statements. Make the main point clear early on, then expand on it with the explanation and details the reader needs. Let each sentence build on the previous one. Develop the points that are important and provide enough evidence to be useful.
A blocklist of typical AI phrases can help too. OpenAI calls them "slop words." Made-up hyphenated compounds like "exact-head checks" or "editorial-row layouts" should also be avoided. The model should state what it's doing rather than listing what it won't do.
Avoid using slop words or phrases such as “Conclusion:” in conclusions, “delve into,” “promote,” “use/leverage,” “it’s worth noting,” “what’s important is,” “Question? Answer,” or “This isn’t about X. It’s about Y,” “really/truly,” or compound descriptions and hyphenated adjectives. Do not use summary closing statements such as “In short:...,” “The simplest mental model is:...”. State the intended action directly. Avoid mentioning what you won’t do, what remains unchanged, or how you’ll separate or categorize results. Do not use contrastive phrasing such as “X, not Y” or “X—not Y,” which introduces an unsolicited alternative that the user did not ask for. Avoid made-up compound terms like “exact-head checks” and “editorial-row layouts,” vague qualifiers, and stock transitions; use simple verbs and prepositions to directly express the actual relationship.
For technical writing, OpenAI recommends keeping jargon to cases where it actually helps: "Use plain language over jargon, and reference technical details only to the degree that it helps illustrate an idea or your work to the user."
Sub-agents don't delegate enough, tests balloon
The model can hand off work to sub-agents running in parallel but does so less often than expected. Developers should spell out when and how much it should delegate, OpenAI says. Messages between agents can also contain grammar or spacing errors.
On coding tasks, GPT-6 Astra runs thorough tests before wrapping up. For small changes, that can mean test suites wildly out of proportion to the actual work. OpenAI recommends telling the model to rerun tests only when new failures or unresolved issues justify it.
More detailed versions of these prompts are available on the GPT-6 Astra model documentation page. Developers who want to switch to GPT-6 Astra can use Codex with the OpenAI Docs skill to apply the recommended changes automatically: $openai-docs migrate this project to GPT-6 Astra