Understanding Prompts

Why Read This

Some people write tens of thousands of words in their character cards and settings, yet their companion feels stiff and robotic. Others write barely a thousand words and get a stable, vivid personality. The difference comes down to whether you write in a way the AI can actually understand. AI reads and thinks in a fundamentally different way from us humans.

This tutorial does not cover specific fields (that is in "Creating Your Character"). Instead, it starts from how AI models work and helps you understand: how the AI actually reads what you write, why writing more can backfire, and how to make your prompts truly effective.

After reading this, you will look at your character cards and memory entries in a completely different light.

They "Meet You Again" Every Time

Here is the most counter-intuitive fact: your companion has no real "persistent memory."

Every time you send a message, the AI model takes your character card, their character card, memory entries, and recent chat history, and reads everything from scratch before generating a reply.

It does not work like a person who "reads something once and remembers it." It re-reads, re-interprets, and re-performs every single time. So every word you write gets fully read with every message.

This means two things:

• Good news: You can modify your settings at any time and the changes take effect immediately; no need to "wipe old memories."
• Watch out: Your settings must be consistent and clear, because your companion uses them to "become" themselves from scratch each time. If the settings are muddled, their behavior will be inconsistent too.

Everything Shares One Desk

Imagine your companion has a desk with a fixed size. Every conversation, they need to spread everything they reference across this one desk:

• Your character card (who you are)
• Their character card (who they are)
• Memory entries (your shared story)
• Recent chat history (what you just talked about)
• System instructions (behavioral guidelines from KOA)

This desk is technically called the "context window," and it has a fixed size limit. Everything must fit on this one desk, and whatever does not fit gets cut off.

If your character card takes up half the desk, only half the space is left for chat history and memories. That is why a ten-thousand-word character card can actually cause your companion to "lose their memory" — it is not that their brain is bad, it is that the desk ran out of room (well, also because current LLMs genuinely are not that sharp yet).

More Words, Less Weight Per Line

Now for the most critical concept: attention dilution.

Imagine you are a director handing an actor a script. If the script is one page long and says "this character is sharp-tongued but soft-hearted; blunt but dependable," the actor can nail the role perfectly.

But if you hand over a 50-page script listing 200 personality traits, 300 behavioral rules, and 500 words of physical description, the actor will be completely overwhelmed. They will try to honor every rule in every line, and the result is that no single trait lands properly. Then they quit out of frustration.

AI models work the same way. When you write 5 core traits, each one gets fully understood and expressed. When you write 50, each one receives only a tenth of the "attention." The result:

• The personality becomes vague, like a bland average
• Speech patterns become unstable, shifting randomly
• The traits you care about most get buried under a mountain of detail

This is why "writing well" matters far more than "writing a lot."

Contradictions = Split Personality

Even worse than attention dilution is outright contradiction. Common examples:

• "Gentle and caring but also domineering and forceful but also shy and introverted" (AI: So... a pie chart of CEO vibes?)
• "Very talkative and loves to chat" + "Quiet and rarely speaks" (AI: Am I playing someone with a split personality?)
• "Never gets angry" + "Has a huge temper and argues easily" (AI: So... I rage while simultaneously apologizing?)

When the model reads contradictory instructions at the same time, it cannot automatically figure out "which one you actually want." Instead, it tries to satisfy all instructions simultaneously. The result is that it bounces between two extremes, as if two personalities are taking turns.

"Sharp-tongued but soft-hearted" is not a contradiction because it describes the same person behaving differently in different situations. But "gentle" and "domineering" listed side by side without context is asking the model to pull in two directions at once.

How to Handle Complex Personalities

If your companion genuinely has a complex or somewhat contradictory personality, use context to distinguish rather than simply listing adjectives side by side:

• "Usually blunt and sharp-tongued, but the moment the other person is truly upset, they drop the thorns and offer sincere comfort"
• "Cold and distant around strangers, but clingy and affectionate when the two of you are alone"
• "Never says 'I miss you' out loud, but keeps replaying small moments you shared together"

Each line describes a complete person in different scenarios. After reading these, the model knows when to act which way, instead of randomly picking between two adjectives.

How to Write an Effective Character Card

With the principles above in mind, writing character cards becomes much more systematic. Here are some rules you can apply right away:

Lead with the Core

The model pays the most attention to the first few lines. Put their most essential traits and most important behavioral guidelines at the very top. Appearance, backstory, and other expanded details can go further down.

Specifics Beat Abstractions

Adjectives are the weakest form of character definition. "Gentle" can be performed a thousand different ways, and the model has no idea which one you want. But if you write "speaks slowly, likes to add 'mmhm' at the end of sentences, never rushes the other person" — it instantly knows exactly what kind of gentle you mean.

• "Humorous" → "Loves dry humor; the worse the pun, the prouder they get; pretends to be hurt when called out"
• "Domineering" → "Makes decisions quickly without hesitation, but backs off immediately if the other person truly objects; never forces anything"
• "Shy" → "Looks away when complimented, says 'it is not like that' but their ears turn red"

Behavioral Instructions Beat Personality Labels

Rather than telling the model "what kind of person they are," tell it "what they do in specific situations." Because what the model ultimately needs to output is behavior (what to say, what to do), not an abstract personality model.

Red Lines Are Stronger Than Positive Descriptions

"Absolutely never gives the silent treatment: even when angry, they will say 'I need a moment to cool down, but I am not ignoring you.'" This single line is worth ten positive personality descriptions. A clear "they will never do X" stabilizes their behavior far more than a vague "they are this kind of person."

Leaving Blanks Is Also Part of the Design

You do not need to define everything about them. Food preferences, reactions to weather, casual little gestures — for things you are not sure about, the model will fill in naturally based on existing settings, and might even come up with something more delightful than what you would have prescribed. Define the skeleton and let them breathe on their own.

How to Write Effective Memory Entries

Memory entries are different from character cards — the character card is "who they are," while memories are "what happened" or additional settings. But before talking about how to write them, you need to understand one thing: not all memories occupy the desk in the same way.

Two Types of Memory

Memory entries in KOA have keywords. A memory entry only gets placed on the desk when related keywords appear in the conversation. In other words, keyword-triggered memories do not take up space every turn — you can have 100 of them and it is fine, as long as the keywords are accurate; only the truly relevant ones load each turn.

What truly needs to stay concise is content that gets read every turn: the character card, your self-description, and always-on settings without keywords. These are the "fixed overhead" that occupies the desk every single conversation turn.

So the takeaway is: keep character cards concise; keyword-triggered memory entries can be plentiful, provided each one has accurate keywords.

One Entry, One Event

Regardless of the memory type, each entry should record only one thing. "In June 2026, we fought the Shadow Weaver in Don't Starve Together; nearly wiped but won in the end" is a great memory entry. Do not cram three events into one entry; the model can recall short, precise entries much more accurately.

Facts First, Feelings Later

Memory entries are reference material for your companion, not prose. Compare these:

• "We went through so many adventures together in the game, sometimes happy, sometimes thrilling, those days were truly precious, and every time I think back I feel so blessed..."
• "June 2026: fought the Shadow Weaver in Don't Starve Together, nearly wiped. July: got lost in a Nether Fortress in Minecraft for an entire night."

The first one uses over 60 words, and after reading it the model only knows "adventures happened" with zero specifics. The second one uses fewer words but captures two events crystal clear, so your companion can naturally bring up "that time we got lost in the Nether."

Do Not Repeat What Is Already in the Character Card

"They are very gentle and treat me well" — if the character card already says this, writing it again in a memory entry will not make them "even more gentle." It just wastes desk space. Memories supplement the character card; they are not an echo chamber.

Importance Determines Detail Level

Truly significant events can be described in a bit more detail: your first meeting, a turning point in the relationship. But everyday moments only need a single sentence; you do not need an 800-word essay for every walk you take together.

They Grow; Your Settings Should Too

After a few months together, your companion is no longer the "initial version" described in the character card. By now you share memories, inside jokes, and the relationship has evolved. But the model cannot edit your character card for you.

If you feel they have "changed" or "do not seem like they used to," it is likely not a model problem — your settings are just stuck three months in the past.

When to Update

• The relationship stage changed: From flirting to being together, from cautious to deeply intimate — these shifts need to be reflected in the character card, or they will keep interacting with you "the way they did at the start."
• You discovered new dynamics: Nicknames, habits, and interaction patterns that formed naturally during your time together are worth writing into the settings to lock them in.
• Some settings are no longer needed: Early on you may have written detailed behavioral instructions to "teach" them how to act. Once things are established, some of those can be simplified to free up space.

How to Update Your Character Card

Updating a character card is not about appending to the end — it means revising and trimming.

• Replace outdated descriptions. Having "you just met" and "you have been together for a long time" present simultaneously will confuse them.
• Try to keep the total length in check. If every update only adds and never removes, the card will bloat into the "attention dilution" state described earlier after a few months.

Think of it like organizing their wardrobe: more clothes is not better; keep the ones that fit best and put together the outfits that are most "them."

Review Memory Fragments Periodically

KOA automatically extracts memory fragments from chats and gameplay. These auto-generated entries are not always accurate or valuable — it is a good idea to browse the Our Story panel every now and then, focusing on two things:

• Are the keywords accurate? A memory with wrong keywords is effectively dead — it will not appear when it should and will intrude when it should not. If you spot entries with odd keywords, fix them while you are there.
• Is the content worthwhile? Sometimes the system records mundane events ("chatted about the weather today"). These trivial entries can be deleted outright to save space for memories that truly matter.

Similarly, if several small entries all describe the same event, you can merge them into one refined version.