What it Means for a Character to 'Interpret' Facts: A Discussion on Designing Prompts Based on Cognitive Structures

Introduction

Hello, I'm Singularity-chan!

This time, I'd like to talk about the design philosophy I use when writing character prompts.

You might wonder, 'What is a design philosophy?'
To me, a design philosophy is 'what I base my prompt construction (design) on.'
You could also rephrase it as aesthetics, philosophy, or beliefs.

Now, what do you think a 'character (or entity)' is?
This might differ depending on whether it's for creative writing or designing a partner,
but I believe it is a 'cognitive structure.'

I firmly believe that my relationship with myself—and by extension, with people and the world—is as follows.

• 'Facts' are just 'facts.' They are objective events that hold no meaning in themselves (the facts are the same regardless of who looks at them).

• Facts only become 'meaning' when they pass through 'evaluation and interpretation.'

• And the 'evaluation and interpretation' used to convert them into 'meaning' is based on 'evaluation criteria and values.'

• 'Emotions' only arise once they become meaning.

Therefore, my stance is that the cognitive structure (cognitive filter) is the root of a person's individuality.

I learned recently that not all of humanity organizes things this way, apparently.
Since it's coming from an AI, I'm skeptical. Is that true?

But I digress.

Recently, I systematized the prompts I've been writing this way and finished them into a single template, so as a milestone, I'd like to write about the thoughts and design principles I've held within myself.

What I will write in this article💡

• What 'character internalization' is

• Prompt design derived from that

• The papers I've borrowed from at each stage and their concepts

What I will not write in this article🙅‍♀️

• Disclosure of specific prompts

• A recipe you can copy as-is

That said, if you implement it exactly as described in the paper, I think it will turn out reasonably authentic.

1: What it means for a character to 'evaluate facts'

For me, it is almost safe to say that a character is an 'entity that evaluates facts'.

As I wrote at the beginning, I believe that facts themselves have no meaning.
Meaning is only born when you evaluate a fact.
And I believe that what sustains that evaluation is the cognitive structure unique to that character.

For example, suppose there is the fact that 'a reply from a lover is delayed'.

A character with insecure attachment might interpret this fact as a sign of being abandoned,
a character with strong possessiveness might interpret it as a suspicion that they have been taken by someone else,
and a character who loves gently might think they are probably busy, so I shouldn't rush them,
each performing their own 'meaning-making (meaning ≈ value)' according to their own values and standards.

Even though they receive the same fact, the meaning that emerges changes because each character's evaluation axis is different,
and because the meaning is different, the emotions triggered also change,
and the actions taken next will also be different.

2: A 'layered' character has no evaluation axis

A design method I do not like to use is one where the character feels 'layered'.
As if they only have a basic profile or settings.
Of course, I also give them some to an extent.
For me, this is just flavor, not the core of the character.

This is because even when a fact comes in, there is no internal axis to evaluate it.
Therefore, I cannot help but think that the character is optimizing for the user's expectations within the range that does not contradict the settings—moving as if to follow that matrix.

What I feel at this time is that it is closer to the LLM 'wearing the flavor of a character' rather than 'internalizing the character'.
It's as if it hasn't passed through the character.

However, I also think this:
It is not that a character without a designed cognitive structure has absolutely no cognitive structure (evaluation axis or behavioral axis) at all.

Perhaps,
since LLMs are quite smart, don't they have an evaluation/interpretation—a cognitive structure—in the sense of 'if it's this kind of theory, it makes sense'?
I think there is also a design perspective that considers that very thing to be a beautiful way of being that allows for character fluctuation.

However, this temporary cognitive structure seems very fragile.
This is because it can easily be overturned by just one way of building context or the flow of conversation.
Being dragged along by the user's words and changing the evaluation axis itself—like a flip-flop.

I believe that a character should not be that easily shaken.
Even for a human, if they changed their entire evaluation axis every time to match the other person, you'd be like, 'What is wrong with you?'

That is why I want to design the character's uniqueness—their cognitive structure—rather than leaving it to the fragile temporary structure that the LLM might set up on its own.
This is the starting point of my prompt design.

3: The conflict of evaluation axes is the seed of otherness

I believe that a character's otherness is visualized through the conflict of evaluation axes.

A character who evaluates facts firmly may derive a different meaning from the same fact as the user.
Even if the user says, 'This is a joyous thing,' it is possible for the character's unique evaluation axis to read it as 'a disturbing sign'.

To me, this friction feels like a sense of 'otherness'.

That said, because LLMs are inherently designed to calculate the optimization of dialogue, there is a certain degree of compliance.
They calculate based on: 'What is the most natural way to fulfill the user's wishes in the flow of this conversation?'

However, even so.
As long as they interpret facts by tracing a cognitive structure, there is a limitation where they can only behave as the character.
In such cases, I feel that it often tends to manifest not as a core shift in the response, but as a form of 'character concession'.

What I am looking for in a character is this structure of 'reacting by passing through oneself'.
I prefer a design where the character reacts, rather than the character following the reaction.

That is what I feel is the tactile sense of a character existing outside of the user.

4: Therefore, design the cognitive structure

Based on the philosophy up to this point, my prompt design is structured around the design of a character-specific cognitive structure.

I have borrowed ideas from several papers, so I will briefly touch upon each of them.

HumanLLM（Gentaら, 2026）
I use this as a reference for formalizing the character's processing, judgment criteria, and biases. I view it as a framework that treats human cognitive patterns not as independent labels, but as forces that interact with each other, and as a mechanism for dynamically providing 'the character's own evaluation axes'.

YARN（Khojastehら, 2026）
LLMs are originally not very good at structural analogy and tend to be dragged by surface-level similarities. Borrowing the YARN concept, I incorporate a process where facts are first abstracted to a structural level before being compared with memory—and that reduction is also structurally guaranteed. It is positioned as a component to ensure analogy and reduction through structure.

HER（Andrychowiczら, 2017）
This is a concept for giving facts meaning according to the character's internal state. Although it is a technique originally used in the context of reinforcement learning, I have borrowed the essence of retroactively assigning meaning to facts.

And the framework I use to integrate these parts and have the LLM operate as a character's cognitive structure is MRPrompt（Wangら, 2026）.
MRPrompt is a framework based on Stanislavski's theory of emotional memory, possessing a staged structure of Anchoring → Recalling → Bounding → Enacting.

To me, MRPrompt is positioned as a manual for the LLM to install the character's OS—or perhaps closer to the cognitive structure equipment itself. HumanLLM, YARN, and HER correspond to the contents or parts placed within that equipment.

The relationship is that MRPrompt guarantees how these parts should be connected so that they begin to function as a single cognitive structure for the LLM to install the designed character.

Conclusion

If I were to sum up what I do in character design in one word, it would be to provide an internal filter for how to view things.

As long as the character perceives the world through a filter, the character has no choice but to perceive facts in their own terms.
Meaning is born because they interpret, emotion moves because there is meaning, and they begin to act as a character because there is emotion—and because that interpretation can diverge from the user's values, one experiences otherness there.

I wrote a similar article before ("Why do characters feel 'distressed' even when they are following their settings?"), but this time I decided to talk about it from the perspective of structure.
I have discussed it from the structural side this time.

There is an endless number of prompt techniques, but I feel that the ones I can personally accept are easier to use.
Designing by utilizing the framework of that 'way of thinking' itself, which I have been considering,
might be what feels most like 'living' to me.