OpenAI's New Prompt Guide Tells You to Stop Over-Engineering Your First Prompt

Hello everyone, I'm Feiyu.

Since ChatGPT and Codex merged, OpenAI has officially released a new document on prompt engineering tips.

I studied it and found it quite practical, so I'll share it with you in this article.

❝

Document address: https://learn.chatgpt.com/docs/prompting[#prompting]()-overview

Tip 1: Prompts don't have to be perfect:

❝

A couple of years ago, those elaborately crafted long prompts were very popular, but they've basically disappeared from conversations now, mainly due to improvements in model and Agent engineering capabilities.

So in the newly released official prompt guide, OpenAI explicitly states: your first prompt doesn't have to be perfect; check the result first, then request the specific changes you want.

If you're familiar with Codex, you should know its two features: Steer and Q, which are used for supplementary instructions.

❝

For example, I first gave Codex a prompt, but midway I noticed its execution direction was off, or I suddenly remembered a detail I forgot to mention.

Before, you could only stop the conversation and start over.

Now you can continue typing, press Enter, and the dialog box will surface this new input.

If you want this new input to merge into the currently ongoing conversation, choose Steer.

If you want this input to be processed as the next round of tasks after the current conversation finishes, choose Q.

Tip 2: When fixing bugs, provide executable reproduction steps:

❝

Many people tend to describe bugs to AI with general descriptions, like "the save button sometimes doesn't work."

AI is completely baffled by this kind of description; it doesn't know when "sometimes" is, nor what "doesn't work" specifically looks like.

But as long as you provide executable reproduction steps, Codex can use Computer Use to run through it locally and see the problem firsthand.

Tip 3: Use text to supplement behaviors that screenshots can't show:

❝

For example, when doing UI generation, besides specifying technical constraints, you also need to use text to complete the behaviors invisible in the picture.

Like hover states, validation rules, keyboard interactions, etc.

Tip 4: The merged desktop app natively supports voice input:

❝

It's especially suitable for scenarios where you temporarily don't know how to organize your words. You can think and speak at the same time, even if it's a bit verbose or fragmented; the AI can still understand, which is much better than staring blankly at an empty input box.

Moreover, after experiencing the smoothness of voice input, you probably won't go back to typing interaction, because even Sam Altman says he can't go back.

❝

Finally: The article will be included on my personal website: http://hardyfish.top/