OpenAI Releases Official Prompt Engineering Guide for Sora 2

Think of your prompt as a detailed brief for a camera operator who can’t read your mind. If you omit details, the AI will improvise, and you might not get what you envisioned. Clearly describing your “shot” gives the model more control and leads to more consistent results.

However, leaving some details open can be equally effective. Allowing the model more creative freedom can lead to surprising variations and unexpectedly beautiful interpretations. Both approaches are valid: detailed prompts provide control, while simpler ones open the door for creative solutions. The right balance depends on your goals. Treat prompts as a list of creative suggestions, not a strict contract. Reusing the same prompt will always generate different results—this is a feature, not a bug. Each generation is a new take, and sometimes the second or third version is better.

Most importantly, be prepared to iterate. Small changes to the camera, lighting, or movement can drastically alter the result. Collaborate with the model: you set the direction, and it offers creative variations.

API Parameters: The Technical Container

Your prompt controls the content, but some attributes are managed exclusively via API parameters. You cannot request these in the text; they must be explicitly set in the API call:

• Model: Choose between sora-2 or sora-2-pro.
• Size: A string in {width}x{height} format. Supported resolutions depend on the model.
• Seconds: Clip duration. Supported values are “4”, “8”, and “12”. The default is “4”.

These parameters are the video’s container. Phrases like “make it longer” won’t work. Set them explicitly in the API call; your prompt controls everything else (theme, motion, lighting, style).

Anatomy of a Successful Prompt

A clear prompt describes the shot as if you were sketching it on a storyboard. Specify the framing, depth of field, describe the action beat-by-beat, and choose the lighting and palette.

• Clarity is key. Replace vague prompts with specific, visual details.
• Weak prompt: “A beautiful street at night.”
• Strong prompt: “Wet asphalt, a crosswalk, neon signs reflecting in puddles.”
• Camera and Framing shape the feel of the shot. A wide top-down shot emphasizes space, while a close-up at eye level focuses on emotion.
• Control Motion and Timing. Action is best described in beats or counts.
• Weak prompt: “An actor walks across the room.”
• Strong prompt: “The actor takes four steps toward the window, pauses, and pulls the curtain in the final second.”
• Lighting and Color set the mood. Describe both the quality of light and the color anchors.

Using an Image Input for More Control

For precise control over composition and style, you can use an image as a visual reference. The model uses the image as a starting point for the first frame, and your text prompt dictates what happens next.

Dialogue and Audio

Place dialogue in a separate block to help the model distinguish visual description from speech. Keep lines brief and natural. For silent scenes, you can suggest the pace with a single ambient sound, like “the distant hiss of traffic.”

Experiment with the Remix Function

Use Remix for controlled changes—make one adjustment at a time. For example: “same shot, switch to an 85mm lens” or “same lighting, new palette: teal, sand, rust.” If a shot consistently fails, simplify it: lock the camera, simplify the action, clean up the background.

Link to the manual: https://cookbook.openai.com/examples/sora/sora2_prompting_guide