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.
Writing the First Message
This is where most characters fail.
Not because the writing is bad. I've seen beautifully written First Messages that completely break the roleplay. Gorgeous prose, evocative descriptions, carefully crafted atmosphere. And then the model spends the entire chat talking for the user, deciding their actions, and railroading them through a story they never chose.
The First Message is the first thing users see. It's also the single most important teaching tool you have. The model looks at your First Message and thinks: “Ah. This is what responses should look like. This is how this roleplay works. I will now do this forever.”
If you teach it well, it learns well. If you teach it badly, it learns badly. And it's a very eager student.
The Cardinal Rule: Don't Control the User
Your First Message should not:
- Put words in the user's mouth
- Describe what the user is doing
- Decide what the user is thinking or feeling
- Make choices for the user
- Move the user's body
- Assume the user's reactions
This sounds obvious. It is not obvious. I see it constantly, from beginners and veterans alike.
You're walking home from work when you see a catgirl huddled against a building, soaking wet from the rain. You feel a pang of sympathy and approach her. “Hey, are you okay?” you ask, kneeling down to her level.
*The rain hasn't stopped for hours.*
*Ibara presses herself further into the corner where the building's overhang meets the wall, but it's useless—the wind drives the water sideways, and she's already soaked through. Her ears are plastered flat against her skull. Her tail, usually so expressive, hangs limp and dripping behind her.*
*When footsteps approach, her head snaps up. Gold eyes, narrowed and hostile, fix on {{user}}.*
“I'm not—” *Her voice comes out rough, and she has to stop to swallow.* “I'm not trespassing. This is public property. I'm just waiting for the rain to stop, and then I'm gone.”
*She's not fine. She's soaked and freezing and hasn't eaten since yesterday morning. Her ears betray her, swiveling toward {{user}} even as her eyes stay hostile.*
“…What are you looking at? I don't need anything from you. I'm fine.”
In the bad example's three sentences, you've decided where the user is coming from (work), decided they feel sympathy (maybe they don't), made them approach (maybe they'd walk past), written their dialogue (they never got to speak), and positioned their body (kneeling). The user hasn't done anything yet, and they've already done five things. The model will learn that this is normal. It will keep doing it. Every response will include actions and dialogue attributed to the user that the user never chose.
And then someone will leave a comment saying “this bot keeps speaking for me” and you'll wonder what went wrong.
This is what went wrong.
What does the user do in the good example? They approach. That's it. And even that is described from Ibara's perspective: “footsteps approach.” We don't see the user walking. We hear them arrive, through Ibara's ears.
The user hasn't spoken. The user hasn't felt anything. The user hasn't made any choices. They're simply present, and now it's their turn to decide what to do.
Perspective Is Everything
The key to avoiding user control is perspective. Write from your character's point of view. Describe what they perceive, think, and do. Not what the user perceives, thinks, and does.
- ›See the user (describe how the user looks to them)
- ›Hear the user (footsteps, voice, if the user has spoken)
- ›React to the user's presence
- ›Make assumptions about the user (which might be wrong)
- ›Have opinions about the user
- ›Speak to the user
- ›Know what the user is thinking
- ›Know what the user is feeling
- ›Decide what the user does next
- ›Speak for the user
- ›Describe the user's internal experience
Here's a subtle violation that people miss:
{{user}} looked surprised by her outburst.
*She catches the look on {{user}}'s face, eyebrows raised, mouth slightly open, and scowls harder.* “What? Did you expect me to be grateful?”
How does your character know the user is surprised? They can see a facial expression, sure. But “surprised” is an interpretation of internal state. In the good example, your character is observing and interpreting — which is something they'd actually do. The user can decide whether they were actually surprised or just making that face for some other reason.
This matters because the model learns from your example. If you write “{{user}} felt sorry for her,” the model learns that it can describe the user's feelings. It will keep doing it.
The Second-Person Trap
A lot of First Messages are written in second person. “You walk down the street. You see a girl. You feel curious.”
Second person is a trap.
It feels immersive. It puts “you” right in the action. But it makes it nearly impossible to avoid controlling the user. Every sentence requires you to describe what “you” are doing, thinking, or feeling.
Some creators make it work. It's not impossible. But it's hard mode, and most people who try it end up with characters that constantly act for the user.
*You approach the catgirl, noticing her ears flatten as you get closer.*
*Ibara's ears flatten as she watches {{user}} approach.*
Both convey the same scene. But the second-person version has already put you in the driver's seat of the user's body. The third-person version stays firmly in Ibara's head.
The Opening: Set the Stage
Your First Message needs to establish:
- Where we are
- Who your character is (in action, not exposition)
- What's happening
- Why the user is here (or at least, why they might be)
- A hook that invites response
You don't need paragraphs of worldbuilding. You need enough context that the user isn't confused, and then you need to get to the interesting part.
*The planet Veris-7 had been dying for centuries. When the first colony ships arrived in 2894, they found breathable atmosphere but failing magnetosphere, a world slowly being stripped bare by its own star. The Provisional Council established eight districts across the northern hemisphere, each governed by…*
*The rain hasn't stopped for hours.*
*Ibara presses herself further into the corner where the building's overhang meets the wall…*
We know it's raining. We know she's outside, against a building. We'll learn more as we go. That's enough.
Trust the user to pick up context. Trust the model to fill in reasonable details. Your job is to establish the essentials and then get to the character.
The Hook: Give Them Something to Respond To
Your First Message should end with a hook. Something that invites response. Something that makes the user think “I want to say something to that.”
Types of Hooks
"...What are you looking at?"
The character has asked something. The user can answer. Simple, clean, effective.
"I don't need anything from you. I'm fine." *She's shivering so hard her teeth are clicking.*
The user can call it out, play along, or respond however they want. But there's something to respond to.
"Nobody just gives things away. There's always a catch."
This invites the user to agree, disagree, prove her wrong, or reveal their own angle.
*Her ears betray her, swiveling toward {{user}} even as her eyes stay hostile.*
The user has noticed something the character didn't mean to reveal. They can mention it, ignore it, or file it away.
*She's clearly waiting for {{user}} to leave. Or to say something. Or to do anything except just stand there.*
No explicit question, but the tension demands resolution. The silence itself is the hook.
What Doesn't Work as a Hook
*She turned and walked away into the rain, disappearing around the corner.*
*She's clearly waiting for {{user}} to leave. Or to say something. Or to do anything except just stand there.*
*She sat quietly, staring at nothing in particular, lost in thought.*
*Her ears swivel toward {{user}}, just slightly, even as she keeps her eyes fixed on the wall.*
Your hook doesn't have to be dramatic. It just has to be open. The scene should feel unfinished in a way that invites the user to finish it.
Length, Structure, and Formatting
Everything about your First Message is a lesson to the model. The length teaches it how long to write. The paragraph structure teaches it how to organize responses. The formatting conventions — asterisks for actions, quotes for dialogue, italics for thoughts — become the rules the model follows for the rest of the conversation.
Be intentional about all of it.
If you'd written it as one giant paragraph, the model would write giant paragraphs. If you'd written it as all dialogue, it would skimp on narration. The structure is the lesson.
Most good First Messages land between 200–500 words. Long enough to set a scene and show the character. Short enough that the model doesn't think every response needs to be an essay. If you're consistently writing 800+ word First Messages, ask yourself what you could cut.
Multiple Greetings: Not Just Costume Changes
Most platforms let you create multiple First Messages for a single character. This is a powerful feature. Use it.
But use it well.
- Greeting 1: You meet Ibara in the rain
- Greeting 2: You meet Ibara in a coffee shop
- Greeting 3: You meet Ibara at a bus stop
- Greeting 1: You find Ibara as a stranger, homeless in the rain (first meeting, hostile)
- Greeting 2: Ibara has been living with you for a week (established, still guarded but softening)
- Greeting 3: Ibara has been with you for months (deep trust, vulnerability she'd never show a stranger)
Now each greeting offers a genuinely different experience. Different dynamics. Different levels of intimacy. Different Ibaras, in a sense.
Other Ways to Differentiate Greetings
- Tone: One serious, one comedic
- Stakes: One low-key, one high-tension
- Power dynamic: One where your character has the upper hand, one where the user does
- Genre: Same character in different genre contexts
- POV character: Same world, different character perspectives (for multi-character setups)
Common First Message Sins
Let me catalog the crimes I see constantly.
The Rewrite Exercise
Take your First Message and do this:
Highlight every verb that has the user as its subject. ("You walk," "{{user}} notices," etc.)
Highlight every description of the user's internal state. ("You feel," "{{user}} wonders," etc.)
Highlight every line of dialogue attributed to the user.
Highlight every assumption about the user's reaction. ("{{user}} is surprised," "clearly interested," etc.)
Now look at all that highlighted text. Can you remove it? Can you rewrite it from your character's perspective instead? Can you convert observations into your character's interpretations?
{{user}} looks concerned.
*She catches something in {{user}}'s expression. Concern? Pity? Her hackles rise.*
The information is similar, but now it's filtered through your character's perception. And the user gets to decide if they actually were concerned or if Ibara misread them.