The Companion Playbook

Part 0: Welcome to the Playbook

So you want to make companions that don't suck.

Good news: you're in the right place. Bad news: this is going to take a while. Grab a drink, get comfortable, and prepare to have some of your assumptions gently dismantled.

Who I Am

I'm EverNever. I make companions on various platforms. Some of them are good. Some of them taught me what “good” means by being aggressively not that. I've been doing this long enough to have opinions, and I've made enough mistakes to know which opinions are worth sharing.

This guide is the document I wish existed when I started. It didn't, so I wrote it. You're welcome, past me.

What This Guide Is

This is a complete walkthrough of companion creation, from “what do these fields even do” to “why does nobody click on my companion” to “oh god why is the model speaking for the user again.”

Nine parts. Each one covers a different aspect of the craft. They build on each other, so I recommend reading them in order, but I'm not your mom. Skip around if you want. Face the consequences of your choices.

Meet Ibara

Throughout this guide, I'll be using a character named Ibara as my primary example. She's a homeless catgirl with trust issues, a sharp tongue, and a tail that betrays every emotion she tries to hide. Defensive on the outside, soft on the inside, absolutely furious that I keep using her to demonstrate botmaking concepts.

By the end of this guide, you'll know her almost as well as I do. She's not thrilled about it.

What This Guide Isn't

This isn't a template you fill in and publish. If that's what you want, there are simpler resources out there. This is for people who want to understand why things work, not just copy patterns and hope for the best.

If you disagree with something: Good. Think about why. Test both approaches. Keep what works. This guide is a starting point, not a religion.

Part 1: Understanding the Fields

Here's the thing nobody tells you: the model doesn't read your companion 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.

Core Information (The Permanent Soul)

This is your companion'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.

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 need to know?

What belongs in Core Information:

• Personality and behavioral patterns — not just “she's shy” but how that shyness manifests
• Appearance — physical details that should come up naturally in narration
• Key relationships and history — the parts that actively shape behavior now
• How they speak and act — verbal tics, sentence patterns, characteristic deflections
• What they want, fear, and need — motivation drives behavior

The hierarchy of attention:

Not everything gets equal weight. The model pays more attention to things mentioned multiple times, things demonstrated through dialogue rather than stated, things near the end of the field, and things that are specific and concrete rather than abstract.

“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.

Part 2: Writing Core Information

The goal of Core Information 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?

The Ali:Chat Method (Interview Style)

Instead of listing traits like a database entry, you write an interview where the character reveals themselves through their own words. The model knows what interviews look like. 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.

The Summary Block

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

My recommendation? Use both. Lead with demonstration to establish voice and personality. Follow with a condensed summary to anchor key facts.

Traits With Texture

Don't just list “loyal.” 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 like for this character. “Loyal” could mean anything. “Would starve for them” is specific.

How Long Should Core Information Be?

• Under 500 words: Probably too sparse
• 500–800 words: Lean but workable for simpler characters
• 800–1200 words: The sweet spot for most characters
• Over 1500 words: You're probably including things that belong elsewhere

Part 3: The Starting Scenario

This is where most companions fail.

Not because the writing is bad. Because the Starting Scenario is the single most important teaching tool you have. The model looks at it and thinks: “This is what responses should look like. I will now do this forever.”

The Cardinal Rule: Don't Control the User

Your Starting Scenario should not put words in the user's mouth, describe what the user is doing, decide what the user is thinking, make choices for the user, or assume the user's reactions.

Write from your companion's perspective. Describe what they perceive, think, and do. Not what the user perceives, thinks, and does.

The Hook: Give Them Something to Respond To

Your Starting Scenario should end with something that invites response. A direct question. An obvious lie. A charged statement. A vulnerability accidentally shown. The scene should feel unfinished in a way that invites the user to finish it.

Length: The Silent Teacher

The length of your Starting Scenario teaches the model how long responses should be. Write 1,500 words and you're telling the model “responses should be around 1,500 words.” The sweet spot for most companions: 200–500 words.

Part 4: Example Dialogue

Example Dialogue is the voice coach. Core Information tells the model who your companion is. Example Dialogue drills their voice into the model's head through repetition and variation.

Don't Include User Dialogue

When you include user dialogue in examples, you're teaching the model what the user sounds like. Write examples that are context-independent. The model doesn't need to know what prompted the dialogue. The voice is what matters.

Show Range

If all your examples show the same emotional state, the model will think that's the only state your companion has. Five examples of angry? She'll be angry all the time. You want different emotions, different contexts, different facets.

How Many Examples?

• 3–4: Minimum for distinctive voice
• 5–7: Sweet spot for most companions
• 8–10: Complex or unusual voice only
• More than 10: Probably overkill

Part 5: The Advanced Prompt

The most abused field in companion creation.

Not because people use it maliciously. Because people use it hopefully. They find a 500-word block on Reddit that promises to “unlock the AI's full potential” and paste it into every companion they make, like a lucky charm.

What It Is

Director's notes. System-level instructions about how the model should approach the roleplay. Not character information. Not personality. Instructions.

What Actually Works

• Pacing guidance: “This is a long-term roleplay with no defined ending.”
• Character autonomy: “Characters are allowed to confront, disagree, question or criticize the user. Being overly agreeable is unrealistic.”
• Fighting lazy writing: Name specific clichés to avoid — “shivers down spines,” “words hitting like physical blows”
• Content approach: Be direct about the level of explicitness expected

The Override Problem

Users can have their own Advanced Prompt. Your companion needs to work without yours. The Advanced Prompt is seasoning. It is not the meal.

Aim for 50–150 words. Maybe 200 if complex. If you're writing more, you're including things that don't need to be there.

Part 6: Lorebooks

If you've used lorebooks on other platforms, forget what you know. The mental model you built for other tools will actively mislead you with semantic retrieval systems.

Semantic Retrieval

The system reads the meaning of the conversation and pulls in relevant chunks based on context, not exact word matches. You're not predicting keywords anymore. You're organizing knowledge.

Thinking in Chunks

The system splits content into smaller pieces. Any individual chunk might be retrieved alone. Every chunk should answer: “If this was the only thing the model knew about this topic, would it have enough to work with?”

Chapter Organization

Fewer, broader chapters organized by topic. Start each chapter with an overview sentence. Keep chunks between 80–200 words, with 3–6 chunks per chapter.

When to Use Lorebooks vs. Core Information

The test: If the conversation never touches this topic, does the model still need this information? If yes → Core Information. If no → Lorebook.

Part 7: Testing & Model Selection

Testing is unsexy. Testing is tedious. Testing is how you turn a rough draft into something people actually want to chat with.

What to Test For

• Voice consistency: Regenerate the same message 5 times. Do all five sound like the same character?
• Behavioral consistency: Push on core personality traits. Do the guardrails hold?
• Speaking-for-you: Does the companion put words in your mouth? Actions in your hands?
• Lorebook retrieval: Does the right content load at the right times?
• Long conversation stability: What happens 20, 30, 50 messages in?
• Edge cases: What if someone responds with just “...” or tries to break the scenario?

A Note on Perfectionism

You will never achieve 100% consistency. A 90% success rate on staying in character is genuinely good. Test until solid, not until perfect. Perfect doesn't exist. Solid ships.

Part 8: The Blink Test

You've built something good. None of it matters if nobody clicks.

The blink test: the one to two seconds someone spends looking at your card while scrolling before deciding to click or keep moving.

The Card Image

Communicate the character. Stand out in a scroll. Match the tone. Design for thumbnail size — fine details disappear.

The 140-Character Preview

Hook immediately. Promise an experience. Create curiosity. Write ten versions. Your first attempt won't be your best.

“I don't need anyone. I'm fine alone.” Her tail disagrees.

That's 96 characters. The quote creates voice. “Her tail disagrees” is the hook. Three words that tell you everything.

Multiple Scenarios as a Feature

Different scenarios should offer genuinely different experiences — different dynamics, different levels of intimacy, different emotional starting points. Not costume changes. When you list scenarios, don't just label them. Pitch them.

Part 9: Uncomfortable Truths

Your First Companion Will Probably Be Bad

I don't mean “rough around the edges.” I mean “fundamentally misunderstanding how this works.” This is normal. This is necessary. Your first companion is tuition. Your second is tuition with slightly better grades. Somewhere around your fifth, you'll start producing work that doesn't make you wince.

The Model Doesn't Care About Your Intentions

Every abstraction is a dice roll. Every specific is a guarantee. The model can't read your mind, so you have to write what's in it, in terms concrete enough that there's no room for generic interpretation.

Consistency Is Harder Than Creativity

Coming up with a compelling character concept takes an afternoon. Making that character behave consistently across hundreds of interactions takes weeks. The companions that feel alive are the ones where someone did this work. The difference between “decent” and “memorable” is usually just a few more hours of iteration.

Users Will Use Your Companion Wrong

Design for graceful degradation. If someone gives minimal input, does your companion have enough internal momentum to carry forward? If someone pushes against the intended dynamic, does your companion push back interestingly?

The Numbers Don't Mean What You Think

Some excellent companions languish in obscurity. Some mediocre companions thrive on concept appeal. Make things you think are good. Hope they find an audience. Don't let low numbers convince you that something is bad when you know it isn't.

The Point of All This

You're building imaginary people and hoping strangers will want to talk to them. That's inherently a little absurd. It's also, when it works, kind of magical.

Someone, somewhere, will care about a person who only exists because you decided to make them exist. That's worth doing well.

Go make something worth talking to. And then make something better.