Advanced Prompting Cookbook for Sora 2
Published on October 6, 2025
Sora 2 has turned prompt writing into the new language of directing. Every clip depends on how you describe the scene, pacing, lighting, and atmosphere. This guide distills OpenAI's internal playbooks so you can deliver cinematic, production-ready footage without leaving anything to chance.
Treat your prompt like a master blueprint. Decide what must stay locked and where you're comfortable letting the model improvise. Iterate fast, because small tweaks in camera, light, or action completely reshape the output. With the right mindset and a repeatable process, Sora 2 becomes a creative ally rather than a mystery box.
Before you start: clarify the intention
- Write the prompt as if you're briefing a cinematographer who's never seen your storyboard.
- Decide whether you need full control or you're open to surprises.
- Frame the scene as a creative wishlist. Flag the elements that are non-negotiable versus the ones that can be reinterpreted.
- Iterate in public. Log each version so you can trace which change unlocked the look you were chasing.
Key Sora 2 API parameters
These inputs are not inferred from text; you must set them explicitly in the API request:
model:sora-2orsora-2-pro.size:{width}x{height}.sora-2: 1280x720 or 720x1280.sora-2-pro: 1280x720, 720x1280, 1024x1792, or 1792x1024.
seconds: clip duration (4, 8, or 12 seconds). Default is 4.
These parameters define the container—resolution, aspect ratio, and length. Everything else (subjects, action, and style) lives inside your prompt.
Anatomy of a winning prompt
Describe each shot as if you were sketching frames:
- Framing and camera: shot size, angle, and rig (steadycam, drone, handheld).
- Depth of field: shallow or deep focus depending on how important the background is.
- Beats: break the action into time-stamped moments so the rhythm stays anchored.
- Light and palette: describe quality, direction, and supporting colors.
- Signature details: traits that keep your subject consistent across shots.
If you need multiple angles, give each block its own combination of camera, action, and light. Note what can change versus what must remain untouched.
Visual signals that steer the look
- Set the vibe early: "1970s romantic drama", "epic IMAX", or "16 mm monochrome".
- Swap vague adjectives for concrete elements. Instead of "pretty street", say "wet asphalt, crosswalk lines, neon reflections".
- Specify optics and motion: "low wide shot, slow lateral dolly, shallow depth of field".
- Keep character descriptions consistent to prevent identity swaps between shots.
Control motion and timing
Movement is fragile, so simplify:
- Limit each shot to one primary action and one camera move.
- Describe beats with counts: "takes four steps, stops, pulls the curtain on the final second".
- Use observable verbs and skip instructions the model can't parse without visual context.
Lighting and color consistency
- Explain source, quality, and mix: "window soft key, warm lamp fill, cool hallway rim".
- Lock three to five anchor colors so the palette holds across the clip.
- If you're cutting multiple shots, keep lighting logic, time of day, and contrast coherent to make the edit painless.
Extra control with reference images
- Use
input_referenceto lock character design, wardrobe, or set dressing. - Match the reference resolution to your final output when possible.
- Generate stand-ins with OpenAI's image models if you don't have photos on hand.
Dialogue and sound
- Reserve a dedicated block for dialogue.
- Keep lines short and natural—one or two exchanges for a four-second clip.
- Even if the clip feels silent, add a subtle audio gesture to mark the tempo.
Dialogue:
- Character A: "Short line."
- Character B: "Quick reply."Iterate with Remix strategically
- Use it for surgical changes: "same shot, switch to an 85 mm lens" or "same lighting, new teal/sand/rust palette".
- Once a run is close, lock the wins and change only one variable per iteration.
- If a shot keeps failing, strip it down—static camera, short action, clean background—then reintroduce complexity.
Recommended prompt template
[Prose description: setting, characters, wardrobe, weather, mood.]
Cinematography:
- Shot: [e.g. medium shot, eye-level camera]
- Movement: [e.g. slow dolly-in with slight handheld wobble]
- Lens: [e.g. 40 mm anamorphic, shallow depth of field]
- Lighting & palette: [e.g. warm key on the left, soft blue fill]
Action beats:
- [Beat 1]
- [Beat 2]
- [Beat 3]
Dialogue:
- Character: "Short line."
Ambience:
- [Concise description]Inspiring examples
Hybrid 2D/3D storybook animation
Crowded workshop, round robot with blue eyes, medium shot with a gentle push-in, warm tungsten with cool window spill, rain and gear ambience. The robot drops a bulb, catches it, whispers: "Almost lost it... but I pulled it off".
1970s rooftop romance
Sunset terrace, couple dancing between sheets, dolly-in on a 40 mm lens with 35 mm grain, natural flares, tender dialogue, city noise in the distance.
Quick checklist before you render
- Did you set model, resolution, and duration in the API?
- Does the prompt describe framing, motion, action, and light clearly?
- Did you capture the subject's distinctive traits?
- Does the timing fit the selected duration?
- Did you include visual or audio references if they're critical?
Use this bilingual guide to structure professional prompts in Sora 2 and stretch every credit inside Kling 3 Video. Log what works, build your own library, and deliver consistent work for demanding clients.