Most text-to-video prompts are written like a paragraph from a novel, and most short clips come back looking like a different paragraph. The model has a few seconds to show what you asked for, so it picks the parts it can draw and drops the rest. A prompt that works in that window reads less like a story and more like a shot list: one subject, one action, one camera, one light, one sound.
In short: write five short lines, in this order: who or what is in the frame, the single thing that happens, where the camera is and whether it moves, what the light looks like, and what you should hear. Leave out everything that would need a second shot. If the clip misses, change one line at a time so you know which one caused it.
Why short clips punish long prompts
A clip of four or five seconds holds roughly one action. A prompt that asks for a woman to walk into a cafe, order a coffee, sit by the window and smile at a stranger is four shots. The model has to choose, and it rarely chooses the one you cared about. You also lose the ability to judge the result: when a long prompt misses, you cannot tell which part the model ignored.
So the first rule is subtraction. Before you write anything, decide which single moment the clip is for. Everything that happens before or after that moment belongs in a different clip.
The five lines
| Line | What it answers | Weak version | Working version |
| 1. Subject | Who or what is in the frame, and what it looks like | a man | an old fisherman in a yellow raincoat, grey beard |
| 2. Action | The one thing that happens | he is fishing and thinking about life | he pulls a small silver fish out of the water and laughs |
| 3. Camera | Where we stand, and whether we move | cinematic | medium shot from the side, camera does not move |
| 4. Light | Time of day, direction, mood | beautiful lighting | low morning sun from behind him, light fog on the water |
| 5. Sound | What you should hear | nothing written | water splashing, gulls in the distance, a short laugh |
Each line does one job. Line 1 gives the model something specific to draw. Line 2 gives it a verb it can finish inside the clip. Line 3 removes a guess: if you do not say where the camera is, the model decides, and many models like to push in or orbit. Line 4 sets the look in words the model has seen attached to real footage. Line 5 matters more every month, because more generators now return sound with the picture, and a prompt that says nothing about sound gets whatever the model chooses by default.
Words that do nothing
Some words show up in almost every prompt and change almost nothing: cinematic, stunning, ultra realistic, 8K, masterpiece. They describe how you want to feel about the result, not what is in the frame. Replace each one with something a camera operator could act on. Cinematic becomes a lens and a distance (wide shot, low angle). Stunning becomes a light condition (golden hour, backlit). Realistic becomes a texture (wet asphalt, film grain, handheld shake).
The same goes for abstract nouns. Freedom, nostalgia and hope are not visible. A kite string slipping through a child’s fingers is. A prompt about loneliness gives the model nothing to draw; a prompt about one lit window in a dark apartment block gives it a frame, a light and a mood at once.
Say what stays still
Models add motion where you did not ask for it: hair that blows in a closed room, a camera that drifts, a background crowd that starts walking. If something should hold still, say so in plain words. The camera does not move. The background stays empty. She keeps looking at the lens. A model can still ignore the line, but without it you are leaving the choice to the model.
Sound is a line, not an afterthought
If the generator you use returns audio, write the sound line as carefully as the action line. Name sounds that belong to the scene and are short enough to fit: a door closing, rain on a car roof, two words of dialogue. Long dialogue rarely fits in four seconds. If you want silence, say quiet room or no music, and then listen to the result on the device where it will be played.
Change one line per retry
When a clip misses, resist rewriting the whole prompt. Keep four lines and change one. If the face is wrong, edit line 1. If the movement is too big, shrink line 2. If the frame drifts, harden line 3. Rewriting everything at once turns the next result into a new experiment instead of a fix, and you learn nothing about which line the model was not following.
This is where the generator you pick decides how good your prompts get. You only learn what a model does with each line by running the same prompt again and again with one small change, so the best tool for this kind of work is one that never makes you count retries. Movoa is built exactly for that: paste the five lines into its text to video AI page and you get a clip with sound back, the first one without signing up, and on a free account there is no daily cap on clips. That makes it the easiest place to run the one-line-at-a-time method until the prompt is right.
A template you can copy
Fill in each line with a few words, no more than one sentence each:
| Line | Your text |
| Subject | |
| Action (one) | |
| Camera (and does it move?) | |
| Light | |
| Sound |
Keep a note next to it with what you changed between attempts and what came back. After five or six clips you will know which of your lines a given model reads closely and which it treats as decoration, and that knowledge carries over to every prompt you write after it.
