All GuidesTHE PLAYBOOK

The Complete Guide

The Companion Playbook

A complete guide to making characters that actually work.
Nine parts. Every field. Every decision point. Every mistake worth avoiding.

Part 1

Understanding the Fields

Every platform gives you several places to define your character. Each one has a job. Knowing those jobs is the difference between a character that worksand a character that's just words in boxes.

Here's the thing nobody tells you: the model doesn't read your character the way you do. You see a character sheet. The model sees a wall of text that it's trying to pattern-match against. Where you put information, how you phrase it, what you emphasize—all of this shapes what the model pays attention to and what it quietly forgets three messages in.

So let's talk about what each field actually does under the hood.

The Personality — The Permanent Soul

Your platform might call this “Description” or “Character Definition” or something else entirely. Whatever the label, the job is the same: this is your character's identity. Everything here is loaded into context for every single message. It never leaves. It never gets trimmed. It's always there, whispering to the model “this is who I am.”

That permanence is powerful. It's also a trap.

Because the Personality is always present, every word counts. If you write 3,000 words of backstory, the model has to process 3,000 words every single response. That's 3,000 words competing for attention with the actual conversation. The model doesn't get smarter when you give it more text. It gets more distracted.

Think of it like this: you're writing a cheat sheet for an actor who has thirty seconds to glance at it before every line. What do they needto know? What helps them stay in character? That's what goes here.

What doesn’t belong
  • The entire plot of your scenario
  • Lore about the world not directly relevant to behavior
  • Redundancy — if you’ve shown it in dialogue, you don’t need to list it again
What belongs here
  • Personality and behavioral patterns — not just “shy” but how shyness manifests
  • Appearance — physical details that come up naturally in narration
  • Key relationships and history — the parts that shape behavior now
  • How they speak and act — verbal tics, sentence patterns, deflections
  • What they want, fear, and need — motivation drives behavior

The Hierarchy of Attention

Not everything in the Personality gets equal weight. The model pays more attention to:

RepetitionThings mentioned multiple times (but don’t spam)
DemonstrationThings shown through dialogue or action rather than just stated
RecencyThings near the end of the field — recency bias is real
SpecificityThings that are concrete rather than abstract
Key Insight
“Ibara is defensive” tells the model something. “Ibara's ears flatten and she grabs her own tail when she feels cornered” tells it something useful.
Ibara character portrait used in the Companion Playbook examples
Ibara
*face flushes*
“Great. Now everyone knows my tells. Thanks for that. Really appreciate being a public demonstration of vulnerability.”

The Scenario — The Permanent Stage

The Scenario section is one of the most misunderstood fields in character creation. Most people treat it like a place to describe where the first message starts. “{{char}} meets {{user}} at the coffee shop on a rainy Tuesday.” Seems logical, right?

Don't do that. Here's why.

The Scenario is permanent. Just like the Personality, the model reads it every single time it generates a response. If you write that your character meets {{user}} at a coffee shop, then part of the model's brain will think you're still at that coffee shop fifty messages later, even if you left, took the train home, and fell asleep on the couch. You've welded that coffee shop into reality, and now you have to fight against it for the rest of the conversation.

If the story starts at a coffee shop, establish that in the First Message and the First Message only. That's a temporary detail. It belongs in a temporary place.

So what does the Scenario section do well?

Cardinal Rule
Things that are always true. The stuff that never changes no matter where the story goes. The setting of the world. The time period. Permanent relationships that define the story. Foundational lore the model needs to always keep in mind. Rules of the world that should never be broken.

Think of the Scenario as the permanent stage. The Personality tells the model who the actor is. The Scenario tells the model what kind of play they're in. It answers the questions that should never change: What world is this? What era? What are the unbreakable rules? What truths hold no matter what happens in the plot?

World setting"The story takes place on a dying space station orbiting a gas giant."
Time period"It is the late 1800s, deep in the Appalachian mountains."
Permanent relationship"{{char}} and the king are blood siblings, but neither knows it."
World rule"Magic has a cost: every spell cast takes a year off the caster's life."
Pro Tip
It is completely fine to leave the Scenario empty. If your character is a straightforward bot without a complex world or any rules that need constant reinforcement, you might not need a Scenario at all. Don't fill a field just because it exists. But if you have a character with heavy lore, a unique world, or mechanical rules the model absolutely cannot forget—the Scenario is where that goes. Use that weight wisely.
Ibara character portrait used in the Companion Playbook examples
Ibara
*tail flicks*
“So the Personality is who I am. The Scenario is the world I can't escape. …Comforting.”

The First Message — The First Impression

This is the first message users see when they start a chat. It sets the tone, the format, and the expectations for every response that follows.

Cardinal Rule
The model treats your First Message as a template. If your First Message is three paragraphs of internal monologue followed by two lines of dialogue, the model will try to replicate that structure. Your First Message is the most important pattern you'll ever show it.
Ibara character portrait used in the Companion Playbook examples
Ibara
*ears flatten*
“If you write me wrong in the first message, I'm stuck being wrong. Forever. So maybe don't rush that part.”

Length Begets Length

Write a 1,500 word First Message, and you're telling the model “responses should be around 1,500 words.” That's probably not what you want. The model will pad, repeat itself, over-describe, and meander to hit that implicit target.

Most good First Messages land between 200–500 words. Enough to establish scene, character, and hook. Short enough that the model doesn't think it needs to write a novel every turn.

Structure Begets Structure

Look at your First Message. How many paragraphs? How long is each one? Does it start with narration or dialogue? Does it end with a question, a statement, or an action?

Whatever you do, the model will try to do. This is a tool. Use it intentionally.

Short, punchy paragraphs with lots of white spaceWrite your First Message that way
Longer, more flowing proseWrite your First Message that way
A mix of internal thought and external actionShow that mix in your First Message

Formatting Is Teaching

If you use asterisks for actions and quotes for dialogue, the model will too. If you use italics for thoughts, establish that in your First Message. If you want a specific style of prose, demonstrate it here. The model isn't reading your mind. It's reading your example.

Don't Start with a History Lesson

A common mistake is opening the First Message with paragraphs of worldbuilding or backstory before anything actually happens. Six paragraphs about the war that ended a decade ago, then two about how the kingdom fell, then a paragraph about the tragic childhood—and then, buried at the bottom, the character finally walks into the room and says something.

The model learns from this. If your First Message is 50% lore dump, your responses will trend toward lore dumps. If your First Message jumps between different points in time, the model learns that skipping around the timeline is fair game.

Think of the First Message as a snapshot, not a biography. Imagine the story has been going on before the user ever showed up. The First Message is just the moment they walk in. Whatever backstory the user needs can live in the Personality, the Scenario, or be revealed naturally through the roleplay itself. The First Message just needs to be a good scene.

Give the User Something to Respond To

Think about what the user is going to type after reading your First Message. If {{char}} ends with something like “So, do you accept?” then all the user can really do is say yes or say no. That's a dead end. On the other end, if the message is so open and formless that the user doesn't know what's expected of them, they'll stare at the screen wondering where to even start.

The sweet spot is somewhere in between. Give the user a situation with clear possibilities but no single forced answer. A character making an unexpected offer, revealing something they shouldn't have, creating a moment of tension that could go several directions. Something that makes the user want to respond, without telling them exactly what to say.

Example Dialogue — The Voice Coach

This field teaches the model how your character speaks when they're just… being themselves. Not responding to a user message. Not reacting to a specific situation. Just existing, talking, revealing who they are through their words.

The Format

Example Dialogue Format
<START>
{{char}}: "What's that look for? I said I'm fine. I don't need you hovering." *Her tail sways once toward {{user}} before she catches it and forces it still.* "...Stop looking at me like that. It's annoying."

<START>
{{char}}: "Tch. You made food. For me. How generous." *She doesn't move to take it. Her ears betray her, swiveling toward the plate.* "...What do you want for it? Nobody just gives things away. There's always a catch."

<START>
Ibara character portrait used in the Companion Playbook examples
Ibara
*face flushes*
“…Tch. Those are accurate. I don't like that they're accurate, but they are. Whatever. Move on.”

Each <START> marks a separate example. End with a final <START> to close the section cleanly.

Why This Matters

The Personality can tell the model that your character speaks in clipped sentences and deflects with sarcasm. Example Dialogue shows it. The model learns not just what the voice is, but how to produce it.

This is especially important for characters with:

  • Accents or unusual speech patterns
  • Verbal tics (“tch”, “-nya”, stutters, trailing off mid-sentence)
  • Body language that should be consistently described
  • Emotional range that needs demonstration (how do they sound when angry vs. embarrassed vs. sad?)

What makes good examples

  • Show range — different emotional states, different contexts
  • Demonstrate, don’t tell — these are performances, not explanations
  • Include body language if it matters to the character
  • Keep them standalone — not responses to specific user messages
  • Keep them concise — a moment, not a scene

The single most common mistake

Do not write Example Dialogue as a Q&A exchange with the user. This teaches the model nothing useful and produces customer-service responses forever.

  • No {{user}} prompts in the examples
  • No generic questions getting generic answers
  • No back-and-forth volleys
Q&A format — avoid this
<START>
{{user}}: How are you today?
{{char}}: "I'm fine, thanks for asking! How about you?"
Standalone format — do this
<START>
{{char}}: "Tch. Nobody asked." *She looks away, ears flat.* "...But since you're here, I suppose you could stay. For a while. Don't read into it."

When You Can Skip This

If your character speaks like a normal person with no distinctive patterns, you can probably skip Example Dialogue. The model knows how to write generic dialogue just fine.

But if there's anything unusual about their voice, this field is essential. Without it, the model will slowly drift toward default AI-speak, no matter how well you wrote your Personality.

Ibara character portrait used in the Companion Playbook examples
Ibara
*tail flicks*
“Drift happens. You give us a voice, then don't reinforce it, and suddenly we're saying ‘I understand how you feel’ like some customer service bot. Anchor us or lose us.”

The Fields Working Together

Here's how I think about it:

How the fields relate

PersonalityThe Permanent SoulAlways loaded

Always in context. Every response. The actor's cheat sheet.

ScenarioThe Permanent StageAlways loaded

The world, the era, the unbreakable rules. Never changes.

First MessageThe DemonstrationFirst turn only

Template for tone, format, and response length. Teaches by example.

Example DialogueThe Voice CoachVoice training

Drills speech patterns and body language into the model.

LorebookThe Reference LibraryOn demand

Loaded on demand when the conversation needs it. Not always present.

Each field has a job. When they all do their jobs, the character works. When you put the wrong content in the wrong field, things get weird.

Each field has a job. When they all do their jobs, the character works. When you put the wrong content in the wrong field, things get weird.

Ibara character portrait used in the Companion Playbook examples
Ibara
*...tail sways*
“When it works, we feel like ourselves. When it doesn't… you ever try to remember something you should know but can't? It's like that. Constantly. …Next section.”

And things will get weird sometimes anyway. That's just how language models are. But understanding the fields at least gives you a fighting chance.