# How to Prompt GPT-6.2 in Codex (and When to Use Astra or Sol)

Use Sol for everyday coding and Astra for harder investigations. Then give Codex a prompt that spells out how you’ll check its work.

Author: Vladislav Guzey
Published: 2026-11-20
Canonical: https://proflead.dev/posts/prompting-gpt-6-2-in-codex/


“Fix the contact form” is the kind of prompt I’d type in a hurry. Here’s the part I’d add before hitting send. The page says “Message sent,” but sometimes the email never arrives. Test what happens with a normal message, one with emoji and line breaks, and one when the email service is down. The third test makes Codex check the case where a visitor sees a success message even though sending failed.

That example is from my GPT-6.2 video. I went through OpenAI’s prompting guide and picked five habits you can use with either model. First, here’s when I’d reach for Sol or Astra. [OpenAI’s GPT-6.2 prompting guide](https://learn.chatgpt.com/guides/prompting-gpt-6-2-astra)

<figure>
  <iframe width="560" height="315" style="width: 100%; height: auto; aspect-ratio: 16 / 9;" src="https://www.youtube.com/embed/auio_b7oj7A" title="GPT-6.2 Astra or Sol? Which to Use and How to Prompt Them" loading="lazy" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" allowfullscreen></iframe>
  <figcaption><a href="https://youtu.be/auio_b7oj7A">Watch my GPT-6.2 prompting tutorial on YouTube</a>.</figcaption>
</figure>

## GPT-6.2 Astra or Sol?

I’d start with **Sol** for everyday work, like reviewing a pull request, updating code after feedback, or making a change with a clear test. I’d use **Astra** when the investigation is the hard part. If messages disappear somewhere between the contact form, a queue, and the email provider, Astra is the model I’d pick to trace where they go. [OpenAI’s GPT-6.2 release notes](https://learn.chatgpt.com/docs/whats-new/gpt-62-astra-sol)

Both models came to Codex, ChatGPT Work, and the API on November 17, 2026. In the Codex app, open the model picker. In the CLI, type `/model`. [OpenAI’s Codex model selection instructions](https://learn.chatgpt.com/docs/whats-new/gpt-62-astra-sol#codex)

## 1. Tell Codex how you’ll check the fix

Describe the bug the way you would to a teammate who already knows the project. Give them the symptom, what should happen, and the steps you’ll use to decide whether the fix worked.

> The contact form sometimes shows “Message sent” when the email never arrives. Keep the layout and validation messages as they are. In a test environment, send a plain message and one with emoji and line breaks. Check that both reach the test inbox. Simulate an email service failure and confirm that the form shows an error and keeps the message available to retry.

Asking Codex to check the inbox matters here. A test that only looks for “Message sent” would pass with the original bug.

## 2. Match reasoning effort to the investigation

- **Medium** for a quick survey. Ask Codex to list the forms that lack error handling.
- **High** when you have a lead to follow across several parts of the system. If the blog got slower after a deploy, ask Codex to check whether the image resizing change explains it.
- **Extra High** when several causes might interact. An intermittent logout could involve cookie settings, caching, and token refresh. Ask Codex to test those possibilities against the behavior it can reproduce.

Give Codex a stopping point at any effort level. For the slow blog, I’d ask for before-and-after load times on the same page under the same conditions. [OpenAI’s guidance on reasoning effort](https://learn.chatgpt.com/guides/prompting-gpt-6-2-astra#match-reasoning-effort-to-the-task)

## 3. Use subagents for independent checks

Before a launch, checking links, reviewing page metadata, and running Lighthouse can happen in parallel. Give Codex the scope and say how you want the results back.

> Check the staging site for broken internal links and missing page titles or meta descriptions. Run Lighthouse on the homepage and one article. Use subagents as needed. Give me one report with the affected URLs, findings, and scores. Leave the files as they are during this review.

If one check needs the results of another, spell out the order. A Lighthouse run meant to measure a fix has to wait until the fix is in place.

## 4. Ask Codex to use the feature in the browser

“Add dark mode” leaves out where the control belongs, whether the setting follows you to another page, and what happens after a reload. Put those steps in the prompt.

> Add a dark mode toggle to the header. At phone width, turn it on, open a blog post, and reload. Confirm that the post stays dark and the toggle is still accessible. Switch back to light mode and reload once more to check that light mode persists too.

If browser access is available, ask Codex to report which steps it ran and what it saw. A saved preference can work while the mobile header still cuts off the toggle.

## 5. Put recurring project rules in AGENTS.md

Save instructions you’d otherwise have to repeat. Tell Codex which command builds the deployable site, where the project defines its approved colors, or what your team means by a pre-launch check. For example:

> When I ask for a pre-launch check, review internal links, page metadata, and Lighthouse scores. Use subagents as needed and return one report before editing files.

That first phrase tells Codex when the instruction applies. Keep the rest focused on conventions it can’t reliably pick up from the repository. [OpenAI’s guidance on AGENTS.md](https://learn.chatgpt.com/guides/prompting-gpt-6-2-astra#keep-your-agentsmd-lean)

## A four-line prompt you can reuse

```text
Change [the behavior].
When I [take an action], I expect [an observable result].
Keep [an existing behavior] working.
Verify it by [the steps I’ll use to check it].
```

Try the last line on a prompt you already use. Write down the clicks, reloads, or inputs you’d use to review the change yourself.

