Corey Knowles

It's quite long.

There's a lot of description about how codecs should behave.

And this is because this is actually the prompt here for GPT-5 model, which are general.

You can use GPT-5 for a bunch of stuff, creative writing all the way to coding.

And so because it doesn't necessarily know by default how to work in a coding harness, we give it quite a lot.

On the other hand, if we go check out the prompt for GPT-5

This model is trained very specifically to work in this harness.

And so the prompt is, well, I don't know if you can tell, but the prompt is much shorter.

And we don't have to tell it anywhere near as much.

And broadly speaking, almost anything that we have in this prompt

Um, we eventually try to delete because the model should, I mean, most things, there's some of some places.

So that's, uh, that's on, uh, on like the prompts in our code and the harness.

And there are some valuable ways to prompt codex.

Like, let's say that every time I prompt, I want codex to work a certain way.

Um, one thing that I can do.

So like, uh,

yeah, one thing that I can do is I can add a file here called agents.md.

And this file is actually a file that a bunch of coding agents support.

So like Codex is just one of them, but also like AMP and Gemini.

And this was like a really cool thing that like when we created this name, agents.md, we intentionally made it generic and like plural to like invite like, hey, like everyone should use this.