Промптопись

Your message will be understood

Don’t be shy—describe the scene in any language, even at a fifth-grader’s level. “I want her to be standing by the window, with rain outside, and the light to be so warm.” That’s it—that’s enough. The AI doesn’t look down on your wording. It doesn’t judge your style. It extracts the meaning—and translates it into its own language. Into a world of light, composition, texture, and mood. And it often produces an image that looks better than what you had in mind. Because you thought in words, but AI thinks in images—and AI has more of them.

So write however you feel comfortable. Clumsily, briefly, chaotically—as long as the idea is alive.

Using examples

The simplest way is to find an image you like and ask the AI to extract anything from it—the color palette, clothing, setting, or background; a pose, lighting, or the whole scene, including every smallest detail. After that, there are two options: you can just use it as is—simply inserting your own character. Or break it down into parts and assemble something new from them. Take the color palette from here, the pose from there, the mood from somewhere else—and now you have a prompt that no one else has.

The second way is through text. Find a prompt that generates a beautiful image. It doesn’t matter what it’s about. What matters is how it’s structured. Then tell the AI one simple thing: “Using this prompt as an example—its structure and the quality of its description—write a prompt and describe my idea.” That’s it. It doesn’t copy the content, but rather the style: the density of the description, the order of the blocks, the tone. And it applies that to your task.

When precision is required

AI, by nature, loves to improvise. In one generation, the lighting might be slightly softer, and in another, the hair might be slightly longer. Usually, this is actually a good thing—the image is always slightly different, and there are some really great results. But sometimes predictability is needed—so that the scene remains virtually the same from one generation to the next.

In that case, the prompt is converted to JSON, a simple “key-value” format. The camera is separate, the lighting is separate, the character is separate, and the environment is separate. The AI interprets this structure differently—not as a stream of text, but as a schema. Each block is in its own place; nothing gets mixed up. The main benefit is hierarchy. You can specify precise parameters, image layers, details, and the placement of those details.

Additionally, a JSON prompt is less susceptible to “Chinese whispers.” The longer a standard prompt is, the more the AI loses track of the beginning by the end. This doesn’t happen with a JSON structure. Each block stands on its own, and its importance doesn’t diminish just because it’s written tenth instead of first. So if you’re running the same scene ten times in a row and want predictability—this is the way to go. You can’t achieve that with plain text; the AI has too much freedom.

Mood, Feeling, and Context

This part of the prompt is usually the first thing people forget. Why should you write “warm nostalgia” or “a scene from an old Japanese movie”? It’s not a technical detail, but it’s often what makes the image come alive.

AI is trained on millions of images; it contains not only a dictionary of objects but also a dictionary of feelings, and for AI, this isn’t just empty words. Behind every such phrase lies a huge number of visual characteristics: specific lighting, color, composition, facial expressions, even film grain. Links to examples work even better. By writing “a shot like from a movie…,” you trigger a prepared set of parameters that the AI has already seen hundreds of times. One phrase, and a dozen details are already in play.

So don’t be shy about describing the mood and feelings that the future shot should evoke. The AI won’t always produce exactly what you had in mind, but it will definitely come closer than if you hadn’t provided any guidance at all. A technical prompt describes what’s in the frame. The mood describes what you feel when you look at it. Without this, even a perfectly composed scene remains just an ordinary picture.

Negative is not always bad

A very clear, simple, and useful part of prompts. The line “negative_prompt: ...” at the end of the prompt is a list of things that shouldn’t be in the frame. The AI is trained on millions of images and doesn’t care that a “phone” specifically shouldn’t be in the frame.

What to write? The basic minimum depends on the scene—for example, plastic-looking skin, overexposure, extra fingers and limbs, blurry hands, watermarks, text, artifacts. Limit the AI’s imagination so it doesn’t veer into cartoony, 3D-rendered, illustrations. Or something specific to the scene: a hat, a smile, panties, a bra, underwear, explicit nudity.

The most important thing that almost everyone gets wrong is to avoid negatives. Don’t write “without a hat,” write “hat.” Don’t write “no smile,” write “smile.” Write about the subject itself. Twenty to thirty relevant items. Place at the end of the prompt. And remember, a negative prompt won't rescue the generation if the scene prompt is badly written. First, you need to balance the main prompt and remove any conflicts, if they exist, and then enhance it with a negative prompt.

The most common mistake

For some reason, most people who write prompts think it's their duty to include not only the necessary text to reference the character's appearance, but also a lot of unnecessary information: eye color, hair color, hair length, ethnicity, “beautiful,” “young,” “gorgeous,” and so on. As if a long-haired blonde with large breasts is the ultimate object.

All of this needs to be removed from the main prompt. Otherwise, the prompt will lose its universality. A good prompt contains no description of appearance at all. The prompt contains only what is happening to the character in that scene.

The reference and the binding prompt already contain information about appearance; there’s no need to drag that information through every new scene. The absence of appearance details makes the same scene prompt suitable for different characters.