OpenAI Shares GPT-6 Astra Prompt Guide: 5 Key Points to Make AI More Responsive and Proactive in Completing Tasks

As GPT-6 Astra is gradually rolled out to ChatGPT paying users, many have likely already started using it. At this point, they’re bound to wonder: compared with the previous GPT-5.6, do prompts need any changes? Earlier, OpenAI added to its developer documentation GPT-6 Astra Prompt GuideThere are 5 key points, and I’ve organized them below for everyone.

How do you write GPT-6 Astra prompts? OpenAI names 5 behaviors that most need adjustment.

1. If the model stops to ask too often, write out the scope of what it can decide on its own.

OpenAI says that while Astra can maintain better coherence in long, multi-step tasks, it is also more likely than GPT-5.6 Sol to stop and ask the user when information is incomplete. If that question would significantly change the outcome, then asking first is a good thing. However, some questions don’t affect the main result, and the model may still stop and wait for you to reply with something like “use your judgment.”

At this point, rather than giving a general instruction like “don’t ask questions,” a better approach is to tell it which situations it can make assumptions on its own and which situations require confirmation. You can refer to the following Chinese translation version:

Work that encourages more autonomy:

你應該根據指示與先前的對話脈絡，推斷使用者的意圖與任務範圍。你的工作是偏向採取行動，並將使用者想要完成的任務執行到底。

當使用者表達想進行新的工作或修正現有問題時，請持續處理，直到使用者想達成的目標完成。自主朝使用者的目標推進（例如視需要建立隔離的 worktree／checkout、解決合併衝突、執行唯讀操作、建立草稿 PR 等），除非這些操作明顯具有破壞性或不可逆。

If your prompt implies authorization, tell the model to follow through to the end:

當使用者的提示顯示他在要求你採取行動，例如「你可以……嗎？」、「我想要……」、「幫我……」及類似表達時，請將其視為要求你執行工作並採取行動的指示。不要只承認你具備這項能力（例如「可以……」）、提出計畫或表示可以繼續。不要為了節省時間、力氣或 Token，而停在無法完整滿足使用者任務的部分成果，或只是「已經夠有幫助」的解法。如果任務需要持續進行，請完成所有必要工作，直到達成預期結果。

Prompt the model to request approval only after it has prepared concrete, reviewable results:

在向使用者提出釐清問題之前，你應先完成依照上下文已獲授權，而且是讓預定操作變得具體、可供檢查所必需的工作。使用者應該要核准的是一個具體、可以檢查的成果。例如在部署變更、寫入外部應用程式、合併 PR 或發佈網站之前，先完成所有必要工作，讓使用者的核准成為最後一步。可逆的任務、唯讀操作、審查或修正，以及先前在對話中已獲授權，或任務指示強烈暗示已經授權的事項，不需要另外取得使用者許可。

The point here isn’t to let AI ignore permissions, but to separate routine judgment from decisions that truly need human input. Things like reading files, organizing data, creating drafts, or carrying out authorized reversible changes can proceed. Deleting data, making payments, publishing publicly, or other actions that are hard to reverse should still be asked about first.

The official also suggested postponing the confirmation point, completing what can be done first, preparing a concrete and verifiable deliverable, and then letting the user decide on the final step.

To get a website live, the model can first complete the pages, review the content, and prepare a preview—then ask for approval only right before actually deploying to the public environment, rather than asking upfront.

2. If skill files and project rules affect each other, explicitly state which one takes priority.

Astra can follow long-form instructions better than the previous generation, but this also makes it more sensitive to skill files, AGENTS.md, and other rules.

If a project contains both “you must ask when data is missing” and “don’t stop to ask questions,” the model has to guess which one takes priority on its own, and in severe cases it may even halt its work entirely.

OpenAI recommends that developers check the skill and rule files the model can access, and explicitly specify priorities in the prompt. For Chinese prompts, you can refer to the following:

使用者的指示優先於技能所提供的指引。如果使用者的明確指示與技能中的指示互相衝突，請優先遵循使用者的指示。

如果某項技能導致你要求許可或確認、暫停工作、讓使用者要求的工作未完成，或偏離使用者的意圖，請說出並連結你實際讀取的那份 `SKILL.md` 檔案，引用相關指示，並簡短說明這項指示如何適用於目前情況。請清楚區分技能中的明確要求，以及你對指引的解讀。

This is especially important for old Codex projects, where rules files have usually accumulated many entries over time. Early on, rules were added to fix a specific mistake, but months later they may conflict with new ways of working. Astra follows rules more strictly, so the impact of old rules naturally becomes more obvious. Rather than patching things up in every conversation, it’s better to periodically remove duplicate, outdated, or contradictory content.

3. If the reply is too long or has too many lists, be more specific about what you want the result to look like.

GPT-6 Astra tends to produce more detailed, well-formatted responses, and may also reuse sentence patterns across conversations. Therefore, if you want it to reply more briefly, simply writing “make it shorter” is often not enough, because the model doesn’t know whether you want to cut the background, examples, explanations, or conclusion.

A more effective way of writing is to directly specify how the article should begin, how much ground each paragraph should cover, when lists are appropriate, and which verbal tics to avoid.

The following is the Chinese translation of the preferred prompt example:

預設使用清楚、簡潔的段落，每個段落只發展一個主要想法。只有在資訊確實彼此並列、具有先後順序，或使用比較方式會更容易理解時，才使用清單；除非無法用清楚的文字表達層級關係，否則避免使用巢狀清單。使用平實、簡單的語言：選擇熟悉的詞彙、具體的例子和精確的動詞。優先使用主動語態和直接陳述。

The following prompts help strike a balance between using clear, coherent language and maintaining professionalism appropriate to the field:

以白話取代術語，技術細節只提到足以協助向使用者說明某個概念或你的工作。以清楚、連貫的方式溝通複雜概念，並依使用者提示詞與上下文所呈現的背景知識程度調整寫法。

For the writing content, if there are words, phrases, or specified formats you don’t want to appear, you can reference the following prompts:

避免使用浮濫的詞語或片語，例如在結論中使用「重點結論：」，以及「深入探討」、「培養」、「善用」、「值得注意的是」、「重要的是」、「問題？答案。」、「這與 X 無關，而是在談 Y。」、「真正地」，或使用帶有連字號的複合描述與形容詞。不要使用「簡而言之：……」、「最簡單的理解方式是：……」這類結論摘要句。

If the content is to be published on a website, you can also add the target audience, estimated word count, heading hierarchy, and citation format at the end. For customer service systems, you can specify that the system should answer the question first, then provide the steps, without adding pleasantries or formulaic reassurances. The more a prompt can describe a visible outcome, the more reliably the model will execute it.

4. If the model isn’t proactive about dividing up work, make it clear when sub-agents should be used.

GPT-6 Astra already has multi-agent collaboration capabilities, but OpenAI notes that it may delegate tasks less proactively than some workflows expect. If your system already provides multiple sub-agents and you merely tell the model to “divide work as appropriate,” it may not know the degree of division you expect.

You can condense the judgment criteria into “whether tasks can be processed in parallel” and “whether division of labor truly saves time or improves quality”:

在任何時候能把任務委派給另一個代理、藉此平行處理工作（無論你是主代理還是子代理），而且這樣做可以節省時間或提升品質，就應該使用協作工具進行委派。

In addition, messages between agents may contain grammar or spacing errors. Using the following prompt can make messages between agents easier to read:

你傳送給其他代理的訊息與最終回答都可能由人類閱讀，因此請確保內容清楚易讀。字詞和／或數字之間一律保留適當空格。

5. Even small changes may require extensive testing, so remember to set stopping conditions for the validation scope.

Astra tends to do full testing when working on code, which is important for major overhauls. But for changes like text edits, adjusting a low-risk setting, or other small reversible tweaks, it can end up taking far more time than the change itself. OpenAI’s recommendation isn’t to drop testing altogether, but to align the level of verification with the risk of the change.

The following is a reference prompt:

不要為可逆、低影響，而且只是照著實作內容寫一遍的變更建立測試。如果你選擇透過測試驗證工作，請確保這些測試具有實質意義，而且確實是驗證實作所必需的。

執行適合本次變更的測試，並完成必要的檢查。這些項目通過後，只有在出現新的變更、失敗或尚未解決的疑慮時，才擴大或重複測試；否則，請繼續把任務完成。

No need to paste all 5 sets of prompts—just handle the one that’s actually having problems.

After reading the above examples, some people might copy everything and cram it into an already long system prompt. That might fix one problem, but new conflicts may appear. So a more practical approach is to first observe where Astra gets stuck in your workflow, then pick a set of rules to add.

Source: KOCPC Chinese