Text to video generator: write a prompt that works

A text to video generator turns a written description into a short video clip, usually between four and fifteen seconds long. You type what the viewer should see, what moves and what the camera does, pick a model, a length and a frame shape, and the model renders the clip from nothing but those words. The hard part is the writing. A vague prompt gets a vague clip, and most people spend their first few tries learning what the model needed to hear. This guide covers how to write that prompt, which kind of model to pick for which shot, what settings matter, and when typing alone is the wrong way to start.

What a text to video generator actually does

Every text-to-video model works from the same inputs:

  • A prompt. The words describing the shot.
  • A length. Most models offer a fixed list of lengths, such as 4, 6 or 8 seconds, rather than any number you like.
  • A frame shape. 9:16 for Reels, Shorts and TikTok, 16:9 for YouTube and websites, 1:1 for a feed post.
  • A resolution. Usually 480p, 720p or 1080p. Higher costs more and takes longer.
  • Sound, on some models. A few models generate ambient sound, effects or even spoken lines along with the picture.

From those, the model invents everything: the subject, the place, the light, the framing and the motion. That freedom is both the appeal and the risk. You get a finished-looking clip from one sentence, and you also get whatever the model guessed for every detail you left out.

How to write a text to video prompt, step by step

1. Describe one shot, not a story

A clip of five or eight seconds holds one shot. “A woman wakes up, makes coffee, drives to work and gives a presentation” asks for four shots and gets a blur of none of them. Pick the single moment you need: “A woman in a grey sweater pours coffee into a white mug in a sunlit kitchen.” If you need the whole story, write it as a list of shots and generate each one separately.

2. Put the subject and the place first

Models give the most weight to what comes first. Lead with who or what is on screen and where they are, then add the details. Name the things that make the look specific: the clothing, the material, the time of day, the weather.

3. Say what moves, and how fast

Motion is the part a still image cannot give you, so spend words on it. “Steam rises slowly from the mug” or “the dog shakes water off its fur, droplets flying” tells the model what to animate. Without a stated motion, many models return a near-still clip with a gentle drift.

4. Say what the camera does

Camera language works well on current models. Useful phrases:

Phrase What you get
Static shot The frame stays still and only the subject moves
Slow push in The camera moves toward the subject
Pull back The camera moves away and reveals more of the scene
Pan left or right The camera turns to follow or reveal
Handheld A slight natural shake, like a phone video
Overhead shot Looking straight down, common for food and products
Close-up The subject fills the frame

Use one camera move per clip. Two competing moves usually produce neither.

5. Name the style and the light

“Shot on film, warm golden hour light” and “clean studio light on a white background” give very different clips from the same subject. One or two style words are enough. A long list of adjectives tends to cancel itself out.

6. Add sound only if the model makes it

On a model that generates sound, describe it in the prompt: “the sound of rain on a window, a kettle starting to whistle.” For spoken lines, keep them short and put them in quotes. On a silent model, sound words are wasted.

7. Generate a few takes, then refine one thing at a time

Even a good prompt gives different clips each time. Generate two or three, keep the best, and change one thing per new take: the motion, the camera or the light. Changing everything at once makes it impossible to tell what helped.

If a clip keeps getting the subject wrong, the prompt is too vague about the subject. If it gets the subject right but the motion wrong, the prompt is too vague about the motion. Fix whichever part failed and leave the rest alone.

A worked example

Here is one prompt rewritten twice, the way it usually goes in practice:

Version Prompt
First try A coffee ad
Better A cup of coffee on a wooden table, steam rising
Usable Close-up of a white ceramic cup of black coffee on a dark wooden table, morning light from a window on the left, steam rising slowly, slow push in, warm and calm, shot on film

The first gives the model nothing to hold on to. The second fixes the subject but leaves the camera, light and pace to chance. The third answers every question the model would otherwise guess at, and it is still only two lines long.

Which kind of model to pick for which shot

Different video models are good at different things, and that matters more than any one being “best”. As of October 2026, the video models in Reel Studio fall roughly into these groups:

Shot you need What to look for Models in the studio that fit
Cinematic shot with sound or dialogue Strong lighting and native sound Veo 3.1, Veo 3.1 Fast
People moving, sport, action Smooth, physical motion Kling 3.0 Pro, Kling 3.0 Standard
Quick drafts and social clips Low price, follows camera moves Seedance 2.0 Fast, Seedance 2.0 Mini
One longer take Long maximum length Seedance 2.5, up to 30 seconds
Cheapest way to test an idea The lowest-cost version of a family Veo 3.1 Lite, Seedance 2.0 Mini

A sensible habit is to draft the shot on a cheaper model until the prompt is right, then render the keeper on a stronger one. The prompt carries over; only the model changes.

When to start from a picture instead

Text-to-video asks the model to invent everything at once. That is fine for a mood shot, B-roll or an idea you are still exploring. It is a poor fit when something must look exactly right:

  • A real product. The model will invent a bottle that looks like yours but is not.
  • The same character across several shots. Each text-only clip draws a new face.
  • A brand look or a specific layout. Words rarely pin down colours and composition precisely.

For those, make or upload a still picture first and animate it, which fixes the look before the motion starts. Our guide to image to video AI covers that route, and how to make an AI video shot by shot shows how to chain shots into a longer piece. For product clips specifically, see the AI product video generator guide.

A good rule: type a shot when any good-looking version will do, and start from a picture when only one version will.

Doing it in one place

You can do all of this on each model’s own app, one subscription per model. The catch is that the best model for a shot changes from shot to shot, and switching between apps means separate accounts, separate balances and different prompt boxes.

Reel Studio puts these models behind one prompt box and one credit balance. You write the prompt once, pick the model, length, frame shape and sound for that shot, and the studio shows the price in credits before you press Generate. If a take fails, its credits come back. The free tier gives 40 credits once, with no card, which is enough to try a few models on pictures and make a first clip; paid plans and one-time credit packs are on the pricing page.

If writing the prompt is the slow part, the free AI video prompt generator turns a rough idea into a structured prompt, and there are model-specific versions for Veo 3 and Kling.

When typing a prompt is enough

You do not need a studio or a plan for every clip:

  • You need one clip, once. A free allowance on any generator will cover it.
  • The shot is generic. Waves, city traffic, clouds and abstract backgrounds come out well from text alone, and stock footage may be faster still.
  • You already use one model and like it. If one model handles all your shots, its own app is the simplest place to stay.

Where one place starts to pay off is when you make clips every week, want the right model for each shot, and would rather see the price before generating than find out from the balance afterwards.

Common mistakes

  • Too many ideas in one clip. One subject, one action, one camera move.
  • Describing feelings instead of things. “Inspiring” means nothing to a model. “Sunrise over a city, a runner crossing a bridge” does.
  • Negative prompts in plain text. Writing “no people” often adds people. Describe the empty scene instead: “an empty street at dawn.”
  • Wrong frame shape. A 16:9 clip cropped to 9:16 loses most of the shot. Pick the shape for where it will be posted before you generate.
  • Judging a prompt on one take. Generate at least two before rewriting.

Make your first clip

Write one shot with a subject, a place, a motion and a camera move, keep it to two or three lines, and generate two takes on a cheaper model. When the prompt gives you the shot you pictured, render it on the model that fits it best. Start free in Reel Studio and try the same prompt on two models side by side to see the difference yourself.

← All articles