How to write prompts for image generation

A prompt is the text description a neural network uses to create a picture. It sounds simple, but this is exactly where the line runs between "wrong again" and "exactly what I wanted". The good news: writing prompts is a skill, not a talent. It breaks down into clear techniques you can pick up in one evening.

The main principle: you dictate, you don't hint

A neural network doesn't guess your intentions or fill in "what's normal". Anything not in the prompt it decides for you — at random. So rule number one: say out loud what matters. Don't like that the background is different every time? Describe the background. Need a specific angle? Name it.

Compare:

• ❌ "a cat" — you'll get a random cat in a random setting.
• ✅ "a ginger tabby cat sits on a wooden windowsill, looking out the window, rain outside, soft evening light, photograph, side view" — you'll get roughly what you pictured.

A formula that almost always works

Keep five blocks in mind. Not all at once, but the more you fill in, the more precise the result:

1. Subject — who/what is the focus. "A young woman in a red coat".
2. Surroundings and action — where and doing what. "walking along an autumn embankment, umbrella in hand".
3. Style and technique — photo, illustration, 3D, anime, oil, watercolour. "cinematic photograph".
4. Light and atmosphere — "golden hour, warm diffused light, light mist".
5. Frame and details — angle, shot size, lens, colours. "medium shot, shallow depth of field".

Build your prompt with this formula and test it right here — see what's still missing for the result to become predictable.

Style decides more than it seems

One word of style changes a picture more than a dozen tweaks to the subject. "A portrait of a girl" as a photograph, anime, oil painting and 3D render is four completely different images. So set the style deliberately and specifically: not just "beautiful", but "in the style of a film still", "like an illustration in a children's book", "studio product photography". A whole chapter is devoted to styles for specific tasks — AI art and styles.

What NOT to do

• A dump of buzzwords. "4k, 8k, ultra detailed, masterpiece, trending" barely helps in modern models and sometimes drags the picture into tasteless "gloss". One precise description beats ten loud adjectives.
• Contradictions. "Minimalist but with lots of detail" — the model gets confused and picks something vague and in-between.
• A prompt that's too long. Past a certain length the model starts to "lose" the beginning. Remove everything that carries no meaning.
• Describing what should NOT be there in plain text. "No people" can, on the contrary, pull people in. For exclusions many services have a separate field (a negative prompt) — use it.

Iteration is the norm, not a failure

Professionals almost never get the final on the first try. The working cycle is: write a prompt → generate 2–4 variants → pick the closest → change one thing in the prompt → repeat. Change one parameter at a time — that way you know what affected the result. Three or four rounds and the picture "ripens".

What's next

You've mastered the main skill. From here the result also depends on the tool: different neural networks "understand" the same prompt differently. Let's work out which to choose for your task.

In the Twelver chat a prompt is written like an ordinary message, and several models are available in one place — handy for comparing how each understands your description.