Why you need a generator
Generation models do exactly what you wrote. «A girl in a cafe» gives an averaged glossy picture, because the model decides everything you left out on its own: shot size, light, angle, style, mood.
A good prompt answers these questions in advance. But to write it, you need to know the language of cinematographers and photographers: what backlight is, a push-in, a wide shot, depth of field. A prompt generator translates your idea into that language.
For video it is also a question of money. Every failed clip generation costs from tens to hundreds of rubles, and a weak prompt multiplies the budget by the number of attempts.
What a good prompt must contain
For an image:
- Subject. Specifically: not «a woman» but «an elderly woman in a knitted sweater».
- Action or pose.
- Surroundings.
- Shot size and angle: close-up, medium shot, wide shot, top view.
- Light: soft window light, backlight, golden hour.
- Style: photograph, 35mm, film grain, illustration, 3D.
- Format: 1:1, 3:4, 9:16, 16:9.
For video, additionally:
- One action per scene.
- One camera move: dolly in, orbit, static, handheld.
- Pace: slow motion, real time.
A detailed breakdown of camera language: how to write prompts for video.
A request template for ChatGPT, Claude or DeepSeek
Copy it and plug in your idea:
You are an experienced cinematographer and prompt engineer for video and image generation AI. I will describe an idea, and you will write a prompt.
Idea: [your idea in your own words] Type: [video or image] Model: [for example, Kling v3, Seedance 2.0, Veo 3.1, Nano Banana Pro] Format: [9:16, 16:9, 1:1, 3:4] Duration: [for video, for example 6 seconds]
Rules: 1. One scene, one action, one camera move. 2. Specify the shot size, angle, light, style and mood. 3. Do not use empty words: «cinematic», «masterpiece», «8K», «ultra-realistic». 4. The prompt in English, 40 to 80 words. 5. After the prompt, give 2 alternative versions with different light and camera movement.
The two alternatives at the end let you compare several approaches on a cheap model and keep the best.
Example: before and after
Before: «a cat on a windowsill in winter».
After: «Medium close-up of a ginger cat sleeping on a wooden windowsill, snow falling softly outside the window, warm soft light from inside the room, cold blue light from outside, shallow depth of field, slow dolly in, calm cozy atmosphere, 35mm film look».
The difference in the result is huge, and it is not about the length of the prompt but about the fact that you made the decisions instead of the model: shot size, light, camera movement, mood.
Where the generator goes wrong
A prompt generator is an AI too, and it has typical weaknesses:
- Too many actions. The generator likes to write «a girl walks in, sits down, opens a laptop and smiles». A video model ruins that. Cut it down to one action.
- Empty words. «Cinematic, epic, masterpiece» often sneak in even when forbidden. Delete them.
- Impossible details. Small text on a sign, an exact number of objects, complex hands in close-up.
- Invented parameters. The generator may write «4K, 60fps» into the prompt, but resolution and frame rate are set in the service’s settings, not in the text. On settings: seed, CFG and steps.
- Real people and brands. If the generator puts a celebrity’s name or a brand into the prompt, the model may refuse. Some models, Seedance for example, strictly filter real faces.
Built-in generators in services
Many generation services have built a prompt generator right into the interface, so you do not have to switch between tabs.
On KADR it is the «Improve» button. It does more than rewrite the text: the AI analyses the task and recommends a model and settings for it, for example suggesting that Kling is better for animating a portrait and Seedance for a story scene. It is free 10 times a day if you have tokens on your balance, and 3 times a day with a zero balance.
For typical tasks KADR also has ready modes where the prompt is assembled automatically: «Product ad» comes up with a script and prompts for an ad video, «Marketplace images» builds infographics from a photo and a list of benefits.
In short
A prompt generator translates your idea into the language generation models understand: shot size, light, camera, style. It can be ChatGPT, Claude or DeepSeek with the right template, or a built-in feature of a service. Check the result: one action per scene, no empty words and no real people’s names. And test the prompt on a cheap model before running an expensive one.
