> ## Documentation Index
> Fetch the complete documentation index at: https://docs.mage.space/llms.txt
> Use this file to discover all available pages before exploring further.

# How to write a prompt on Mage

> The prompt structure that works on every Mage model: subject, setting, light, and framing, plus how to call Characters and References with @ handles.

A prompt is the text description you generate from. One clear sentence beats a paragraph of adjectives, and the structure below works on every model Mage runs.

## What a prompt needs

| Element | What it decides                   | Example                         |
| ------- | --------------------------------- | ------------------------------- |
| Subject | Who or what the image is of       | a mountain climber in her 40s   |
| Setting | Where they are                    | on a ridge above the cloud line |
| Light   | The mood, and most of the realism | warm backlight at dawn          |
| Framing | How close, and through what lens  | waist-up portrait, 35mm         |

```text theme={null}
a mountain climber in her 40s on a ridge above the cloud line, warm backlight at dawn, waist-up portrait, 35mm
```

That is a complete prompt. Light and framing are the two people skip, and they are the two that separate a generated-looking image from a photographic one.

## Calling a Character or a Reference

If you have built a Character or a Reference, call it with an @ and its name. The rest of the prompt works normally around it.

```text theme={null}
@sasha wearing @redcoat, standing in @cafe, evening light, 50mm
```

That is how you hold the same face, outfit, or location across a whole series. See [Characters](/features/characters) and [References](/features/references).

## Tag-style prompts

Anime models such as Anima and the Illustrious and Pony checkpoints also take comma-separated tags. On Anima, escape literal parentheses as ( and ), because parentheses otherwise read as prompt weighting. Pony checkpoints respond to score tags at the front of the prompt.

## Making a result repeatable

When a generation works, save the prompt and the seed together. Those two plus the model and its settings are what let you return to a look weeks later, or produce the variation a client asked for.

## Changing one thing at a time

The fastest way to lose a good result is to rewrite the whole prompt at once. Change the light, or the lens, or the model, and regenerate. If the base model is the problem, more words will not fix it; switch models. See [which model should I use](/models/which-model-should-i-use).

## What prompting cannot fix

Resolution is set by the model and your plan. Asking for 8K in the prompt does not raise the output resolution; the Image Enhancer app does. Hands, text, and fine repeated detail remain the common failure points across every model, so check those at 100% before you use an image commercially.