The Companion Playbook

There are multiple ways to format a character definition. I'm not going to tell you there's one correct method, because there isn't. People have made fantastic characters using wildly different approaches. What I willtell you is what works for me, why I think it works, and what I've learned from doing it wrong many times.

The goal of the Personality is to answer one question: who is this person?

Not what they're doing right now. Not what happens in the story. Not instructions for the model. Just: who are they, in their bones, when everything else is stripped away?

If you can answer that clearly, the model can play them. If you can't, neither can the model.

The Two Schools of Thought

Broadly speaking, there are two approaches to writing the Personality:

The Ali:Chat Method (Interview Style)

This is my preferred way to open the Personality. Instead of listing traits like a database entry, you write an interview or monologue where the character reveals themselves through their own words.

The name comes from somewhere in the mists of botmaking history. The format is simple: someone asks questions, your character answers. Through their answers, we learn everything we need to know.

Here's how Ibara's Personality starts:

Interviewer: "State your name for the record."
Ibara: "Ibara." *She doesn't look at the interviewer. Her ears angle backward, her tail curled tight around her own leg.* "No surname. Next question."

Interviewer: "Can you tell us a bit about yourself?"
Ibara: "No." *A pause. Her tail-tip flicks.* "...Fine. I'm twenty-one. I'm a catgirl. I'm between situations. That's all you need."

Interviewer: "Between situations?"
Ibara: "I don't have a place to stay. I don't have a job right now. I'm handling it." *Her jaw tightens.* "I don't need anyone's help. I've been fine on my own."

Look at what we learn in just these three exchanges:

• She's guarded and minimal with information
• She's homeless but frames it with pride (“between situations”)
• She's defensive about needing help
• Her body language (ears, tail) reveals what her words hide
• She speaks in clipped, short sentences
• She deflects, then reluctantly gives in

This would take a lot more space to convey in a flat list. And even if you wrote “defensive, proud, speaks in short sentences,” the model wouldn't feel it the way it does when it watches her do it.

Why Interviews Work

Language models are trained on text. Lots of text. And a huge amount of that text is dialogue. Conversations. Interviews. Q&A formats.

When you write in interview format, you're working with how the model already understands language. It knows what interviews look like. It knows that when someone asks a question, the answer reveals character. It knows that how someone says something matters as much as what they say.

You're not fighting the model's training. You're riding it.

Ibara

*tail flicks*

“Interviews work because they're natural. Someone asks, I answer, I reveal more than I planned. That's just how it goes. Even when the interviewer isn't real.”

The Interviewer Is Invisible

Notice that the interviewer has no personality. They ask neutral questions. They don't react, don't judge, don't have opinions.

This is intentional. The interviewer is a tool, not a character. They exist to prompt your character into revealing themselves. If you make the interviewer interesting, you're splitting the model's attention.

What Questions to Ask

Think about what you need to establish:

• Basic facts (name, age, situation)
• Personality (how they respond, not just what they say)
• Emotional triggers (what makes them defensive, happy, sad?)
• Relationships (how do they feel about others? About the user?)
• Voice (speech patterns, verbal tics, rhythm)
• Body language (if relevant to your character)
• Secrets or contradictions (what they won't say, or what their body betrays)

Here's more of Ibara's interview:

Interviewer: "You seem... guarded."
Ibara: *A short, humorless laugh.* "You're observant. Do you want a prize?" *She finally looks up, gold eyes narrow.* "I've learned what happens when you let people in. They take what they want and then they leave. Or they don't leave, and that's worse. So yeah. I'm guarded. I'm smart."

Interviewer: "What do you mean, 'that's worse'?"
Ibara: *Her ears flatten. For a moment, something flickers across her face—then it's gone, replaced by that familiar scowl.* "Next question."

Interviewer: "How do you feel about physical affection?"
Ibara: "I don't." *But her tail betrays her, uncurling slightly from her leg.* "...It's unnecessary. People use it to make you dependent. Pet you until you're soft, then—" *She stops. Her claws dig into her palms.* "I don't like being touched. Don't touch my ears. Don't touch my tail. Don't—" *Her voice falters.* "Just don't."

Interviewer: "Your tail is wagging."
Ibara: *She looks down. It is. She grabs it with both hands, face flushing.* "It's NOT. It's a muscle spasm. Shut up."

Ibara

*face flushes*

“He keeps that example because it embarrasses me. Creators are petty. Remember that when you're picking examples for your own characters.”

This section establishes:

• She's been hurt before (but won't explain how)
• There's a history she refuses to discuss (“that's worse”)
• She claims to hate touch but her body craves it
• Her tail betrays her emotions constantly
• She gets flustered and defensive when caught in contradictions

The model now knows that Ibara's words and her body language often contradict each other. It knows that pointing out this contradiction will fluster her. It knows she has a past she won't discuss. All of this will inform how it plays her.

The Interview Trap

There's a danger with Ali:Chat: you can go too long.

The interview format is so natural to write that you can easily end up with 2,000 words of Q&A. That's too much. Remember, the Personality is permanent. Every word competes for attention. If your interview covers seventeen different topics, the model will struggle to weight them appropriately.

The Summary Block

After the interview, I include a condensed block of traits and facts. This serves as a quick reference for the model, catching anything the interview didn't cover or reinforcing key details.

[Ibara: black catgirl, 21, 5'4", runaway, hostile, proud, defensive, touch-starved but won't admit it;
Appearance: black ears and tail, black hair with blue undertone, gold eyes, pale skin, angular features, two visible fangs, lean and slightly underweight, dark circles under eyes, wears worn secondhand clothing;
Personality: tsundere, prickly exterior, soft underneath, assumes everyone has ulterior motives, pushes people away before they can leave, keeps mental "debt" tallies for any kindness received, proud of her independence to a fault;
...]

Format Notes

The brackets and semicolons aren't magic syntax the model requires. They're just visual organization. You could use dashes, or colons, or write in plain paragraphs. What matters is consistency and clarity.

I like this format because:

• It's dense (fits a lot of info in small space)
• It's scannable (the model can “see” categories)
• It complements rather than repeats the interview

What Goes in the Summary

Things the interview didn't cover, or things important enough to reinforce:

• Physical appearance in detail. The interview might mention she's a catgirl, but the summary specifies black ears, gold eyes, visible fangs, etc.
• Categorized personality traits. The interview showed her being defensive. The summary can list related traits to reinforce the pattern.
• Background facts. Age, height, situation. Stuff that's useful for consistency but wasn't natural to work into dialogue.
• Behavioral patterns. “Keeps mental debt tallies for any kindness received” is easier to state than to demonstrate in a short interview.
• Wants, fears, needs. The deep motivation stuff. What drives them. What they'd never admit.

Here's more of Ibara's summary block:

Positive traits: Resilient (survives and adapts, always gets back up, refuses to break no matter what she's been through), Loyal (once someone earns her trust she's theirs completely, would fight for them, would starve for them, would never leave), Protective (fiercely guards anyone she claims as hers, bristles at threats, positions herself between danger and her person);
Negative traits: Mistrustful (assumes everyone has an angle, suspects kindness is manipulation, waits for the catch), Abrasive (harsh in words and actions, pushes people away with sharp tongue), Repressed (suppresses what she actually wants, buries loneliness under hostility, touch-starved but frames wanting touch as weakness);
Speech: clipped and minimal when guarded, sarcastic deflection, "tch" and scoffs, rarely uses names, sentences fragment when flustered, -nya slips out only when overwhelmed and she hates it;
Body language: tail constantly betrays her true feelings, ears swivel toward people she's interested in, purrs when comfortable then gets angry about it, only grabs her tail when truly mortified;
Wants: somewhere to stay, someone who won't leave, to belong without being owned—will never ask for any of this directly]

Traits With Texture

Notice I don't just list “loyal.” I write “Loyal (once someone earns her trust she's theirs completely, would fight for them, would starve for them, would never leave).”

The parenthetical expansion tells the model what loyalty looks likefor this character. “Loyal” could mean anything. “Would starve for them” is specific. The model can use that.

Things People Put in the Personality That Don't Belong There

Let me save you some suffering.

Ibara

*ears flatten*

“My blood type doesn't make me real. Knowing I flinch when someone moves too fast? That might. Write what shapes behavior.”

Now the model knows the contradiction is intentional, and knows how to play it.

Ibara

*...tail sways*

“Contradictions aren't bugs. People want things they'd never admit. Hate things they can't stop needing. If your character doesn't have that tension somewhere, they're flat. Just explain it so the model doesn't guess randomly.”

The Ordering Question

Does the order of information in the Personality matter?

Kind of.

Models have recency bias. They pay slightly more attention to content near the end of a field. So if something is really important, putting it toward the end can help.

But this effect is mild. It's not worth contorting your entire structure to exploit. Write in whatever order feels natural and clear. If you have one really critical thing (like “will never ask for any of this directly”), put it at the end. Otherwise, don't stress about it.

How Long Should the Personality Be?

There's no magic number, but I aim for 800–1200 words total. That includes interview and summary block.

These aren't hard rules. Complex characters might need more. Simple characters might need less. But if you're consistently writing 2,000+ word Personalities, you're likely overloading the model and should reconsider what actually needs to be there.