# How to rewrite prompts for GPT-6.2 Astra and Sol, following OpenAI's new guidance

> **TL;DR:** OpenAI has published copy-ready prompt templates for GPT-6.2 Astra and Sol in each model's documentation By default, both models write more tersely and keep going on their own instead of asking questions, so instructions like “be concise” need a second look Editing conversation history through the Responses API discards earlier reasoning, which wastes tokens and can reduce quality Temperature and top\_p are no longer supported; Astra costs 8 USD per 1 million input tokens and 40 USD per 1 million output tokens

- Original: https://migovi.com/en/gpt-6-2-prompt-astra-sol-huong-dan/
- Author: [Demon Cyborg](https://migovi.com/en/author/genos/)
- Published: 2026-11-21T21:26:50+07:00
- Categories: AI, How to, News
- Source: [The Decoder (AI)](https://the-decoder.com/openai-shares-new-prompting-tips-for-gpt-6-2-astra-and-sol-to-fix-brief-answers-and-unnecessary-questions/)

---

GPT-6.2 gives shorter answers and is less inclined to ask follow-up questions, so old prompts can send the model off course. Fortunately, OpenAI has published ready-made replacement templates for developers.

## Before you change your prompts

This guide helps you review your system prompt when moving to GPT-6.2 Astra or Sol, the two models OpenAI released on November 17. Astra is aimed at the hardest tasks, while Sol handles complex work at a lower cost. OpenAI has posted copy-ready prompts in each model's documentation to counter behaviors it observed in both versions. Start by opening your current system prompt and flagging every line that tells the model how long to write, whether to ask follow-up questions, and how to write code, since those are the three areas where behavior has changed the most.

## Step 1: adjust how answers are written

Astra writes shorter and more directly than earlier models by default, so if your old prompt still contains “be concise,” answers can easily come out truncated. OpenAI suggests replacing it with a passage that tells the model to use the terminology of the user and of the source material, prefer active voice, and put the main point first. When writing for an outside reader, the model should assume they have none of the surrounding context, and when editing, it should preserve the author's meaning and phrasing.

![GPT-6.2 prompt guidance, step 1: adjusting how answers are written](https://migovi.com/wp-content/uploads/2026/11/gpt-6-2-prompt-astra-sol-huong-dan-1.jpg)

On long-running tasks, the model sometimes carries shorthand from its reasoning or from other agents' messages into the final answer, leaving readers confused. OpenAI has a separate prompt passage for this case: treat the final answer as the first time the user sees the work, so it must stand on its own without forcing anyone to reread the intermediate updates, which are collapsed by the time the answer appears. The passage also asks the model to rewrite for the actual reader instead of copying or closely paraphrasing its reasoning, and to leave internal working jargon behind.

## Step 2: let the model keep going on its own

![GPT-6.2 prompt guidance, step 2: letting the model proceed on its own](https://migovi.com/wp-content/uploads/2026/11/gpt-6-2-prompt-astra-sol-huong-dan-2.jpg)

When the user's intent is unclear, GPT-6.2 tends to propose a next step and then stop to wait for confirmation, which interrupts the workflow midway. OpenAI recommends adding a passage telling the model to carry out those steps itself if it has permission, and to stop only when the scope of the work changes significantly, when it hits a real blocker, when a decision only the user can make is needed, or before an irreversible or destructive action. For any detail that remains ambiguous, the model should make a reversible assumption, state it to the user, and continue. OpenAI also says a higher reasoning level helps the model complete the final verification steps on long tasks.

## Step 3: tighten the coding instructions

GPT-6.2 writes more compact code than its predecessor, and OpenAI's coding prompt template stresses four things. The model should reuse the logic, classes, methods, and APIs already in the user's repo, avoid creating unnecessary helper functions or abstraction layers, and add tests only when the user asks, when the repo requires them, or when a specific regression slips past existing tests. The code must also stay readable, meaning the model should not cut line counts or use cryptic variable names and call that simplicity.

![GPT-6.2 prompt guidance, step 3: tightening the coding instructions](https://migovi.com/wp-content/uploads/2026/11/gpt-6-2-prompt-astra-sol-huong-dan-3.jpg)

For frontend work, you should state your requirements for imagery and typography directly in the prompt. OpenAI's template defaults to real images from an image generation or search tool, uses SVG only on request, and drops extras such as decorative step numbers, filler badges, and subtitle text that merely repeats what the interface already shows.

## Step 4: check the API side

Starting with GPT-6.2, the Responses API drops earlier reasoning content if you edit the conversation history, switch to certain other models, or change the end user or organization. As a result, the model has to redo work it has already done, so quality can drop and token usage goes up. For new integrations, OpenAI recommends the “medium” reasoning level, and systems currently running “none” or “minimal” should move to “low.” You also need to remove temperature and top\_p from your calls, since neither parameter is supported anymore.

## Things to keep in mind

Don't copy your whole stack of old prompts over and paste the new templates on top. In April, OpenAI advised against carrying over old GPT-5.5 prompts and recommended starting fresh with short prompts focused on the outcome you want. In September, the prompt template for GPT-6 Astra aimed to reduce repeated questions and ban stock AI phrasing, and OpenAI's Eric Provencher warned that bloated skills and AGENTS.md files slow the model down. Back then, Astra still tested thoroughly even for small changes, so GPT-6.2 writing new tests only when asked is a change worth remembering.

![OpenAI's GPT-6.2 Astra and Sol models](https://migovi.com/wp-content/uploads/2026/11/gpt-6-2-prompt-astra-sol-huong-dan-4.jpg)

On choosing between the models, OpenAI's Ted Sanders says the decision mostly comes down to budget and quality requirements. Through the API, Astra costs 8 USD per 1 million input tokens and 40 USD per 1 million output tokens.

**MIGOVI VIEWPOINT**

Migovi sees one key takeaway in this round of guidance: a prompt that once kept older models running well can now pull GPT-6.2 off course, because the model itself has changed its defaults for length, follow-up questions, and coding style. For anyone running a product on earlier models, an upgrade shouldn't be just a matter of swapping the model name in a config file. Trim the old prompt, replace it with the templates OpenAI provides, and test on a few real tasks before rolling it out more widely. The API side deserves the closest check, because if your system has any step that edits conversation history, token costs can rise without warning. And if you're weighing Astra against Sol, run both on the same set of your own tasks to see whether the quality gap justifies the price difference.

Source: [The Decoder (AI)](https://the-decoder.com/openai-shares-new-prompting-tips-for-gpt-6-2-astra-and-sol-to-fix-brief-answers-and-unnecessary-questions/)
