How to write a prompt for AI video

A good AI video starts with a clear description. This guide covers what to write, what to skip and how long to go. It ends with three real prompts from the DV4 Showcase and the clips they made.

How DV4 reads your request

You can write one line. A planning model reads it first and writes a shot list, with a prompt for each shot. DV4 films each shot, adds sound, music and text, and hands you one finished video. This means:

Spoken lines go in quotation marks. Write the words you want a person to say inside quotes. The people on screen say them, with their lips in sync. DV4 copies each line into its shot prompt word for word.
On-screen text is set by DV4, not drawn by the video model. Every shot prompt tells the model to draw no text. Titles, callouts and captions are laid out over the footage in a separate step.
The planner may change your request. When it alters something, it lists the change so you can see it.

Dialogue and music are included on Lite, Plus, Pro and Team. See pricing for what each plan includes.

What to describe

Lightricks, the company behind the LTX models that film DV4's shots, publishes a prompting guide. Its advice comes down to six parts. Each is something you can add to your request.

The shot. Name the framing with film terms: a close-up, a medium two-shot, a wide shot. Say whether the camera is still.
The setting and light. Describe the place, the light source, the colors and the mood. Use one consistent light logic per shot. Mixed light sources confuse the result.
The action. Write what happens, in order, from start to finish. Use verbs that move: turns, smiles, looks up. A description of how things look gives the model nothing to animate.
The people. Give age, hair, clothing and anything that sets them apart. Show feelings through the body. "A faint, knowing smile" works better than "she feels amused."
The camera. If the camera moves, say how and when. Say where the subject ends up after the move.
The sound. Name the ambient sound, the music and any speech or singing. Put dialogue in quotation marks. Give the language or accent if it matters.

What to leave out

Lettering in the scene. The LTX guide says the model still handles on-screen text unevenly, and exact spelling is not guaranteed. DV4 sets text itself, so ask for it as a title or callout instead.
Crowds of characters. A few clear people and actions read better than a busy frame.
Chaotic physics. Lightricks notes that very chaotic motion can add artifacts. Simple, plausible motion is more reliable.
Tag lists from other tools. The LTX guide recommends one flowing paragraph over tags or numbered shot lists.

How long to write

For one continuous shot, Lightricks suggests about four to eight descriptive sentences, in present tense. Close-ups need more detail than wide shots. A longer scene can run longer, as long as every sentence adds something you can see or hear.

Pace the action in the words. LTX times a clip to the action you write. It won't add a pause you didn't ask for. If you want a beat of silence, write "she pauses."

Speech also takes time. DV4 checks that a shot's spoken lines fit its length, at no more than 2.5 words per second. A 10-second shot holds about 25 spoken words. Longer dialogue needs a longer video or more shots.

Worked examples from the Showcase

These prompts were filmed on LTX-2.5 and are copied word for word.

Sixty Years

Photorealistic. A static medium two-shot at seated eye level frames an elderly couple in their eighties side by side on a wooden porch swing at sunset: on the left a man with white hair, a lined face and a plaid flannel shirt, on the right a woman with short silver curls, reading glasses and a lavender cardigan. Warm low sunlight from camera right falls across their faces, and the garden behind them is softly out of focus. They sit still and relaxed. The woman turns her head slightly toward him with a faint, knowing smile and says in a soft, dry, conversational voice, "You forgot our anniversary again, didn't you? Sixty years, Harold." He holds her gaze for a moment, a small sheepish smile forms, and he answers quietly in a low, warm voice, "Forgot? Never. I was saving it for a surprise, my dear." She shakes her head gently and smiles. The camera stays locked. Both performances are understated and natural, the voices gentle and close. After the last line nobody speaks again: every mouth stays closed, and only soft evening crickets are heard until the end.

What it does well: it places each person on the left or the right, so the model knows who speaks. Each line has its own voice: soft and dry for her, low and warm for him. The last sentence says what the sound does after the dialogue ends.

Six Hours for This View

Photorealistic. A static medium close-up at eye level frames a woman in her thirties with long dark hair and a heavy olive raincoat, standing still on a grassy sea cliff under a stormy grey sky. Flat, cool overcast light; the sea below is softly out of focus. A strong wind from camera left blows strands of hair across her face and ripples her collar. She squints into the wind with a quiet, happy smile, then calls out over the wind in a warm, breathless voice, "This is it. I drove six hours for this view, and it is exactly where I wanted to be." She takes a slow breath and keeps smiling at the sea. The camera stays locked. Her performance is joyful but natural, her voice raised only enough to carry over the wind. After the last line she does not speak again, and only the roaring wind and crashing waves are heard until the end.

What it does well: the wind has a direction and a visible effect on her hair and collar. The voice matches the setting: raised just enough to carry over the wind. The light stays one thing, flat and overcast.

Moonlight on the River

Photorealistic. A static medium close-up at eye level frames a jazz singer in her forties with close-cropped natural hair, gold hoop earrings and a deep red satin dress, standing at a vintage chrome microphone on a small stage in a dim club. A warm amber spotlight comes from camera right with a soft blue rim light behind her; the club tables are dark, blurred shapes. A slow jazz ballad plays on piano and upright bass. She closes her eyes and sings softly in a rich, smoky alto, slow and legato, "Moonlight on the river, carry me home tonight, carry me home." Her face stays relaxed and tender as she sings. After the line she opens her eyes and gives a small, closed-mouth smile while the piano continues. The camera stays locked. Her performance is intimate and restrained, the vocal gentle and close to the microphone. After the last line she does not sing again, and only the slow piano and upright bass play until the end.

What it does well: it names the instruments and the tempo of the music. The sung line is in quotes, with a voice type and a way of singing. Two lights are named, each with a place and a color.

Common mistakes

Describing a mood with no action. Give each sentence something that happens.
Leaving dialogue out of quotation marks. Without quotes, DV4 can't tell which words are spoken.
Writing more speech than the video can hold.
Asking the scene to show a sign or a slogan. Ask DV4 for a title instead.
Ending with no instruction for the sound. Say what plays after the last line.

Start with the shot and the action, then add light, people and sound. See more results on the Showcase, or read which AI video models film each shot.