You're using Sora 2 wrong. Here's how to be ahead of 99% of users

Sora 2 prompting guide

OpenAI just released a prompting guide for Sora 2.

While I find it useful, it feels like it’s written for API users and video experts. To bridge that gap, I put together this guide to explain everything in plain English.

Right off the bat, you should know that:

• Generating videos isn’t an exact science, but the prompting tips in this guide will improve your chances of getting the results you want

• Be prepared to iterate. Running the same prompt multiple times will produce different outputs, so even a solid prompt can occasionally miss the mark

As a rule of thumb:

• Detailed prompts: Gives you control and consistency. The model will try to follow your guidance, but might not always do so reliably

• Lighter prompts: Gives the model more creative freedom. You can expect surprising results

Both types of prompts are fine!

In fact, while writing this article, I tried some ultra-detailed prompts and got unexpected results. When choosing your prompt style, focus on what your specific clip needs.

This guide includes: prompt tips, a prompt template, and invite codes to Sora 2. The invite codes are at the end of this article, and this time it’ll be for my paid subscribers only (sadly, not everyone shares their own codes after using mine, and some paying readers still haven’t gotten in)

#1 Control motion

If you’ve been creating AI videos for some time, you know movement is the hardest part to get right. OpenAI recommends keeping it simple and describing action in beats or counts: small steps, gestures, or pauses.

Here’s an example of a weak prompt:

Actor walks across the room

And here’s an example of a strong prompt:

Actor takes four steps to the window, pauses, and pulls the curtain in the final second.

Here are more examples (weak prompt → strong prompt):

Cyclist moves quickly → Cyclist pedals three times, brakes, and stops at crosswalk

Dog jumps on a sofa → A dog pads to the sofa, pauses, sniffs the cushion. Hops up, lands softly on the sofa, curls tight, and rests.

#2 Dialogue and Audio

OpenAI recommends putting any dialogue in a block below your prose description so the model clearly distinguishes visual description from spoken lines.

For dialogues, keep this in mind:

• Limit your lines so the timing can match your clip length

• 4-second shot will usually accommodate one or two short exchanges. 8-second clips can support a few more. Long, complex speeches are unlikely to sync well

• For multi-character scenes, label speakers consistently and use alternating turns

<Video description>

Dialogue:

- Detective: “You’re lying. I can hear it in your silence.”
- Suspect: “Or maybe I’m just tired of talking.”
- Detective: “Either way, you’ll talk before the night’s over.”

• For silent shots, you can still add background sounds like “distant traffic hiss” or “a crisp snap”

A starbucks employee is preparing coffee. There’s no dialogue

Baclground sound: The hum of espresso machines and the murmur of voices form the background.

Note: If you have Sora 2 pro, you’ll be able to control video resolution and length

#3 Control visual cues

When it comes to visual cues, style and clarity are important:

• Style: Guides the model toward a desired outcome. “1970s film“ will set a visual tone that frames all other choices in the video

• Be clear: Instead of “moves quickly,” be specific and say “jogs three steps and stops at the curb.”

When you want to cut multiple clips together, keeping lighting logic consistent is what makes the edit seamless.

Here’s an example:

Camera shot: wide shot, low angle
Depth of field: shallow (sharp on subject, blurred background)
Lighting + palette: warm backlight with soft rim

A man stands in the desert, sharp against a blurred RV. Warm backlight rims his hat and shoulders; dust glows, shadows lengthen as he lifts gun, breath steady.

In the following section, I’ll share with you a Sora 2 prompt template, a prompt example, and some invite codes I just got today.