Text-to-Video Prompt Structure: Subject, Motion, Camera, Guards

Text-to-video fails for a boring reason: the prompt describes a picture instead of describing a shot. "A woman in a red coat in a snowy street, cinematic, 8k" tells the model what should be in frame but nothing about what happens, where the camera is, or what must not change.

The fix is grammar, not adjectives. Four parts, always in the same order, and the text-to-video guide explains the shots this grammar plugs into.

The four parts

1. Subject. Who or what, in one sentence, with the details that must survive the shot: age range, clothing, material, colour.
2. Motion. One action, described as body parts and direction.
3. Camera. One move: static, push-in, pull-back, lateral pan, orbit, handheld follow.
4. Guards. What must not appear: no cuts, no text, no extra people, no scene change.

Then two optional lines: light (soft window light, hard noon sun, neon at night) and pace (slow, real-time, slight slow motion).

Before and after

Vague:

Cinematic shot of a barista making coffee, beautiful lighting, high detail.

Structured:

Subject: a barista in a dark green apron and rolled sleeves, short dark hair, working behind a wooden counter. Motion: they press the portafilter down with one hand and set a white cup under the espresso spout. Camera: static medium shot at chest height, slight handheld drift. Light: warm morning light from a window on the left, soft shadows. Guards: no cuts, no text, no customers, hands stay anatomically correct.

The second prompt is longer but every sentence is an instruction. The first is a wish.

Rewrite habits that pay off

• Name body parts, not feelings. "Shoulders rise and fall" reads better than "she looks nervous".
• Pick one camera move. Two moves in one prompt usually produce a slow zoom that also drifts.
• Keep the subject description short and stable. Long costume descriptions drift mid-clip; two or three fixed details hold better.
• Put guards last. The model weights the end of the prompt heavily on what to avoid.
• Drop quality words. "4K, masterpiece, trending on ArtStation" changes nothing except length.

Choose duration before you write

• Four to six seconds covers a single action and a single camera move.
• Eight to ten seconds fits a two-beat action, such as walk-in then turn.
• Longer than ten needs a real reason: a slow orbit, a reveal, a build. Otherwise the second half invents motion you did not ask for.

Short clips also cost less, and the workspace prices every job before it runs.

Choose the aspect ratio for the destination

• 9:16 for reels and shorts.
• 16:9 for a landing page or a pre-roll.
• 21:9 for a wide hero band.
• 1:1 where both orientations get cropped.

Write the prompt after you pick the frame. A "wide establishing shot" in 9:16 is a different shot than in 21:9, and the model will not tell you it had to compromise.

When text-to-video is the wrong tool

• You already have the photo. Use image-to-video. A supplied still locks identity and composition for free; a text prompt has to invent both.
• The shot depends on a real product. Use a product photo with a camera move; text-to-video will redesign your packaging.
• The message needs readable text. Video models still garble letters. Put text on top in an editor.
• You need the same character in five clips. Start from reference images — see the reference-to-video workflow.

Prompt lab

Edit a tested prompt before you generate

Pick a direction, change the details to match your source, then copy it or open the workspace with the prompt already loaded.

Cinematic establishing shotProduct revealEveryday loop

Wide frame, slow camera, one subject.

Editable promptA cinematic establishing shot at golden hour: a lone traveler walks along a coastal cliff path. Slow lateral camera move, warm rim light, soft haze, realistic lens flare. One subject, natural motion, no cuts, no text, no logos.

Open workspace with this promptCopy prompt

Browsing and copying prompts is free. Generation uses purchased credits, and the workspace shows the credit cost before you run a job.

A test loop that costs little

1. Generate a four-second version of the shot.
2. Check the subject first, then the motion, then the camera, then the guards.
3. Change exactly one of the four parts.
4. Only extend to eight seconds when the four-second version passes.

Most prompts fail at part one. If the person in the clip is not the person you described, do not touch the camera line — rewrite the subject and try again.

Frequently asked questions

How many words should a text-to-video prompt be?

Sixty to a hundred and twenty words covers all four parts with one example of light. Longer prompts rarely add control; they usually add contradictions.

Can I use the same prompt on a different model?

The structure transfers, the results do not. Keep the prompt, re-test the duration and the guards on any new model before shipping.

Does text-to-video produce audio?

The workspace has a generate-sound toggle, on by default, and it depends on the provider supporting it. Treat any generated audio as a scratch track: check it before publishing, and add music or voice in an editor when the result matters.