Sora Prompt From an Existing Video: What Transfers and What Doesn't

Sora prompting is closer to writing a shot list for a camera crew than to typing keywords, and that makes it an interesting destination for a reference video: someone has already directed the shot, and your job is to write down what they directed. Here is the conversion, plus the parts of a reference no extractor can hand you.

What Sora-style prompting favours

• Scene first, subject second. Establish the place, then what happens in it — the model builds space unusually well.
• Shot progression in words. “Begins wide and settles into a close-up” is a direction it can follow; “cool camera work” is not.
• Physics and material. Liquids, fabric, dust, reflections and weight read as the differentiator for this family.
• Coherence across the clip — a single sentence describing what stays constant matters as much as what changes.
• Plain, declarative prose. No keyword soup, no negative lists, no weighted syntax. Write it the way you would brief a DP.

Text-to-video models differ in exactly these habits, which is why a prompt written for one family underperforms on another without a rewrite.

The conversion, step by step

1. Extract the structure. Reference clip into SocialToPrompt — a URL from any of 27 platforms, or an uploaded MP4, MOV or WebM, which skips the extraction credit — and take back subject, camera, light, pacing and style. A full run on a clip up to 20 seconds costs 4 credits per the site’s pricing; the free 3-credit signup allowance does not quite cover one, as that arithmetic explains.

2. Reorder for space. Lead with the environment, then the subject inside it, then the camera’s behaviour toward them.

3. Convert fields into sentences. Delete the comma-separated parameter style; Sora reads prose better than tag lists, so the same information becomes two or three clean sentences.

4. Add the physical cause. Where the extraction says “steam rises”, write why it moves: curling and dissipating in the light shaft.

5. Declare duration and ending, which no extractor ever supplies.

What extraction cannot tell you

Three gaps that need a human on the reference clip, in order of how much they hurt the render:

1. Intent and continuity. What the shot is for — the reason the camera waits — is invisible in pixels.
2. Sound and dialogue, if your generation includes audio: a visual extractor cannot hear the clip. Read the transcript where one exists (our caption guide covers pulling it as TXT or SRT).
3. What the camera does between samples. Fast pans and brief whip moves fall between frames, so short, energetic references often come back described as calmer than they were.

Worked example

Reference: 9s travel clip — drone over a coastal road, descends toward a red car, ends on roof.

Extracted fields: coastal road, cliff, ocean, red car driving, drone descending, bright daylight,
                  vivid colours, fast then slow, ends on the car.

Sora-shaped prompt: A two-lane coastal road carved into a green cliff above a flat grey ocean,
late morning and clear. A single red car comes into view from the far end and drives toward the
camera. The shot begins as a high aerial view of the whole curve, then descends steadily, the
cliff edge sweeping past underneath, until the car fills the frame and the road becomes a strip
of grey asphalt beside it. The water stays flat and colourless behind. Nine seconds, ending on
the car's roof as it passes out of frame.

Notice what changed: parameter language became spatial sentences, the descent was written as a progression with intermediate stages, and the ending got stated. That is roughly a minute of editing per clip.

Access, rights and honesty

Two practical notes. Access to Sora-class generation sits behind OpenAI’s own products and API terms, so a prompt written here is portable but the endpoint is someone else’s policy — read it before building a workflow on it. And the rights position on a reference is unchanged by the model you use: describing a shot mechanism is fine, reproducing someone’s creative expression, talent or brand marks is not, and the transformation test is the one that matters. For the model-agnostic version of the conversion, see field-by-field conversion; for other destinations, Wan, Hailuo and Runway.

Frequently asked questions

Can I create a Sora prompt from a video I found?

Yes: extract the clip's structural fields with a video-to-prompt tool, then rewrite them as plain declarative prose that leads with the environment, describes the camera's progression through the shot, and states the duration and ending. Sora responds to briefed sentences far better than to keyword lists.

What does an extractor miss for a Sora-style prompt?

Three things: the intent behind the framing, any audio or dialogue, and fast camera movement that falls between sampled frames. Those need a human who watched the reference, which is why the rewrite step is not optional.

Why should I convert a parameter-style prompt into sentences?

Because prompt dialects differ. Comma-separated slot lists work well for models trained on short production captions; Sora-class models are stronger when the description establishes space and then moves through it, which reads as prose rather than tags.

Is using someone's video as a reference a copyright problem?

Reusing the mechanism — framing, camera progression, pacing, lighting pattern — is not what copyright protects; reproducing their expression, talent, script or brand marks is. Change subject, setting and palette, and ask whether the original creator would recognise their specific work in yours.