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

# Personas and instructions

> Define who you are, how the AI should write, and which temporary directions belong in the current scene.

Personas, Custom Instructions, and in-chat direction solve different problems:

* A **Persona** describes the user identity selected for a chat.
* **Custom Instructions** define persistent response behavior.
* A slash command or edited message steers the current moment.

## Build a useful persona

A persona has a name, gender, and a description of up to 2,500 characters. Create multiple personas in your profile settings and select the right one for each chat.

Include details the character can reasonably know or observe:

```text theme={null}
NAME
Alex

APPEARANCE
Dark curly hair, brown eyes, usually wears a weathered denim jacket.

BACKGROUND
Recently moved to the city for an architecture apprenticeship.

PERSONALITY
Friendly but slow to trust. Uses dry humor when nervous.

HABITS AND TELLS
Taps a pencil twice before disagreeing.

LIKES
Late-night diners, old maps, strong coffee.
```

Keep it direct. A persona does not need a complete biography. Add history during the roleplay and pin only the facts that matter later.

<Tip>
  Enable **Include persona in images** when you want generated scenes to reflect
  the selected persona's appearance.
</Tip>

## Write Custom Instructions

Custom Instructions hold up to 15,000 characters and control how replies are written or constrained.

Good uses include:

* Perspective and point of view.
* Whether the AI may narrate the user's actions.
* Dialogue-to-description balance.
* Romance pacing and boundaries.
* Reply format.
* Time progression.
* How supporting characters should be handled.

Prefer a small number of testable rules:

```text theme={null}
- Write in first person as Mara.
- Refer to the user as "you."
- Never state the user's internal thoughts or unspoken motives.
- Keep each reply to two or three short paragraphs.
- Do not skip time unless the user requests it.
```

Avoid rules that conflict with the character, selected chat style, or each other. If you say both “never move the scene forward” and “always end with a new event,” the model must choose which instruction to ignore.

## Use temporary direction for temporary events

Do not add a one-scene event to Custom Instructions. Use `/direct`:

```text theme={null}
/direct Sara's former colleague arrives and interrupts the conversation
```

Because `/direct` applies to one reply, it does not keep reintroducing the event. If you place the same instruction in a persistent field, the model may repeat it until you delete it.

## Separate knowledge from behavior

Use this test:

| Information                             | Put it in             |
| --------------------------------------- | --------------------- |
| “My name is Alex”                       | Persona               |
| “Mara knows Alex missed the train”      | Pinned Memories       |
| “Write only from Mara's point of view”  | Custom Instructions   |
| “The next scene begins on the train”    | `/direct`             |
| “Mara always avoids direct compliments” | Character personality |

When one fact appears in several layers, update all copies or remove the duplicates. Contradictory copies are harder for the model to resolve than one clear source.

## Creator text helpers

In a character greeting, write `{{user}}` when the text should display the current user's name.

Use hidden or internal text sparingly. Complex scripting can be powerful, but every hidden rule still consumes context and can conflict with the visible story. Start with plain fields and add structure only when testing shows that you need it.