How to Write AI Prompts

A prompt is the text description of what you want the AI to generate. Almost everything depends on it: the same model returns wildly different work for «a cat» and for a detailed description. Below — what a working description is built from, how a weak prompt differs from a strong one, and how to turn one idea into a dozen different shots by changing a single block.

You don't have to write it yourself

If you'd rather skip the theory, describe the idea in a couple of words and the prompt generator will expand it into a full description: «cat in a spacesuit», «visa photo», «birthday card for a colleague». The AI returns a ready prompt to copy or send straight to generation.

The same «Develop idea» button sits inside every tool — there it also accounts for the aspect ratio and style you picked.

Open the prompt generator

What a working prompt is made of

The model cannot guess what you have in mind. The more specific the description, the smaller the gap between what you imagined and what you get. A working structure has five blocks:

Subject — who or what is in frame: «a ginger cat», «a woman in her thirties», «a ceramic mug»
Action and pose — what is happening: «sitting on a windowsill», «looking into the camera», «standing on a wooden table»
Setting — where it happens: «a Scandinavian kitchen», «an autumn park», «a plain white studio backdrop»
Light and mood — «soft morning light», «backlit sunset», «cold neon» — light changes the feel of a shot more than anything else
Style and technique — «photorealistic», «watercolour», «oil painting», «studio shot on an 85mm lens» — without this the model picks a style on its own

Weak prompt

a cat

Same idea, spelled out

A ginger cat sits on a windowsill looking outside, rain running down the glass, soft diffused light of an overcast morning, warm tones, photorealistic, shallow depth of field

A weak prompt and a strong one: what changed

Take one task — «I want a picture of a cat» — and watch what makes a description work. Each step adds exactly one block from the formula above.

Subject only
a cat

The model invents everything else: breed, angle, background, style. The result is unpredictable — none of ten attempts will match what you had in mind.
+ action and setting
a ginger cat sits on a windowsill looking outside, rain running down the glass

Now there is a scene. The shot is recognisable, but light, time of day and drawing manner are still the model's call.
+ light and mood
a ginger cat sits on a windowsill looking outside, rain running down the glass, soft diffused light of an overcast morning, warm cosy tones

Light sets the emotion more than any other block: the same scene in a backlit sunset or in cold neon reads completely differently.
+ style and technique
a ginger cat sits on a windowsill looking outside, rain running down the glass, soft diffused light of an overcast morning, warm cosy tones, photorealistic, shallow depth of field

A finished prompt. Without this last block the model picks a manner on its own, and identical requests return a photo one time and an illustration the next.

Generate from this prompt

Variations: one idea, ten different shots

Keep a finished prompt as a base and change one block at a time. That turns a single description into a series rather than random pictures: you always know exactly what changed.

Base

A ginger cat sits on a windowsill looking outside, rain running down the glass

Generate from this prompt

Light and time of day

soft diffused light of an overcast morning, warm tones
backlit sunset, long shadows, golden haze
night, cold neon signs from the street, blue glints on the fur

Style

photorealistic, shallow depth of field
watercolour, soft bleeds, paper texture
hand-drawn animation still, bold lines, saturated colours

Angle and framing

close-up of the face, eyes in focus
wide shot of the room, the cat silhouetted against the window
low angle from the floor, the cat looming larger

Mood

cosy and calm, a quiet evening at home
melancholy, solitude, muted colours
curiosity: the cat crouches, tracking a bird beyond the glass

Change one block at a time. Change three at once and you won't know which one improved or ruined the shot — and you won't be able to get back to the good version.

Three mistakes that break a prompt

Negations — «no people in the background», «not cartoonish» — the model often latches onto the word itself and draws exactly what you forbade. Describe positively instead: «an empty street», «photorealistic».
A 300-word wall — the longer the description, the more requirements the model drops along the way. One or two dense sentences beat a paragraph of listings.
Magic words — «8k», «masterpiece», «best quality» helped older model versions; modern ones barely react. Spend that space on specifics instead.

What to change when the result misses

AI is a tool, not a «make it beautiful» button: some attempts go sideways even for experienced users. It is almost always one of five reasons, and each has its own fix.

Doesn't look like the source photo — faces transfer far more accurately from a front-facing shot in even light, without glasses, hats or heavy shadows. Group photos, profiles and distant shots drift the most — try a different photo.
Artifacts and broken details — extra fingers, odd objects in the background and melted text are ordinary generation noise, not a failure. Every attempt differs, so a repeat with the same prompt often fixes the frame by itself.
Generated something else entirely — the model took your wording literally. Drop negations («without people» often reads as «people»), describe positively what should be there, and add the missing details — who, where, wearing what.
Blurry or low quality — check the quality tier: pricier modes run stronger models and hold more detail. A finished image can also be upscaled to 4K separately.
Wrong style — if you don't name a style, the model picks one. Say it outright — «watercolour», «oil painting», «studio photo», «flat vector illustration» — or attach a reference to follow.

Quality modes and when each one matters

Quality tiers differ not in «prettiness» but in the model behind them and the cost per generation. The standard mode fits drafts and idea hunting: fast and cheap, and an idea needs several attempts anyway. Higher tiers pay off once the concept is settled and you need the final frame — they hold fine detail, on-image text and multi-object scenes better.

Photos with faces are a separate story. There the source matters more than the tier: a good front-facing shot in the standard mode almost always beats a dark profile photo run at the most expensive setting.

Frequently asked questions about prompts

What is a prompt in simple terms?

A prompt is a text brief for the AI. You write what you want to see, and the model generates an image, video or text from that description. The more specific it is, the closer the result lands to what you pictured.

Do I need a prompt for the ready-made tools?

No. In tools like greeting cards, caricatures or photoshoots the prompt is built in: you upload a photo and pick the options, and the service writes the scene description itself. Prompts matter when you generate an image from scratch or want to depart from the preset.

Which language should I write the prompt in?

Write in your own language — the description is translated and expanded automatically before it reaches the model. There is no need to translate prompts into English yourself.

How long should a prompt be?

One or two specific sentences usually suffice: subject, setting, light, style. Very long 300-word descriptions work worse — the model starts dropping some of the requirements.

Why does the same prompt produce different images?

That is how generative models work: every run starts from random noise, so two runs of identical text give different frames. It is useful — you can regenerate without touching the description until a good one comes up.

Can I skip writing prompts altogether?

Yes, that is what the prompt generator is for: describe the idea in a couple of words and the AI expands it into a full description. It costs 1 energy point.