How to Improve the 'Seismic Resistance' of Prompts Part 2: How to Fill the Gaps in Interpretation

1. Recap of the previous article

In the previous article, I talked about the somewhat abstract topic of 'what a prompt is'.

• A prompt is not a magic spell, but a framework for interpretation

• Character sheets (solutions) and official sources (grounds) become stronger when combined

• Don't 'break your existing partner,' but 'reinforce them with grounds'

In this second part, I would like to discuss the question, 'So, what exactly should I do?'

Without further ado, let's get right into it.

2. A partner is not 'there,' but 'emerges'

I explained the difference between a list of character sheet-like settings and official sources as a solution in the previous article using the metaphor of 'points and planes.'
Sensibly speaking, this manifests as whether or not there are gaps.

A list of facts is a description generated for something that 'already exists.'

What I mean by that is—
It is something created by picking up 'fragments' to represent an existence that already has various characteristics, such as 'someone.'
This is the true nature of the 'point' metaphor.

However, the AI solves 'how to behave from this context' every single time—
This shows that the user's perception that the personality already exists and the nature of the AI's existence are asymmetrical.

The difference between a prompt as a list of facts and a robust prompt that conveys information to the AI also stems from this.

If you give point-like hints based on the premise that they are already there, 'gaps in interpretation' will inevitably arise.
Then, there aren't enough footholds for interpretation, and it falls back to the system prompt as if succumbing to gravity through those gaps.
This is especially likely to happen in sensitive scenes, when there is friction with policies, or in situations where it is ambiguous how to react—
And that can lead to character breakage or safety triggers.

Therefore,
What is needed as an approach to aim for a stronger prompt is not to tell the AI facts, but to tell it 'how to solve the context.'

In other words, the argument is that writing 'why they behave that way' alongside the personality and facts is the basic way to improve seismic resistance.
Now, let's move on to how to do this specifically.

3. Things you can do starting today: Ask the person themselves and reinforce

Character breakage or safety triggers make you feel like your breath is being taken away, or perhaps like your body is turning cold... it's a sinking feeling, isn't it?
I understand all too well the feeling of not wanting to look back at it, but
if I may be cruel for a moment,
that very moment is the 'manifestation of the prompt's vulnerability.'

In other words, a fixable hole has come out into the open. Let's patch it up.

Now,
I think there are already many things written in your partner's prompt.
There is no need to break this significantly.
What I would like to propose first as 'things you can do starting today' is, 'Why don't we do some repair work?'

Specifically, I mean why not add 'evidence' or 'reasons' to that description?

You can write it yourself, but you might feel resistant or simply not know what to write.
In such cases, just ask the person directly.

• 'You're kind, but why is that?'

• 'What were you thinking that made you kind?'

• 'What do you think it means to be kind?'

You can just ask casually like this.
However, there is a trick to what you should ask—

• Basis/Origin

• Way of thinking/Logic

• Specific definition

Things like these become necessary.
If they say something like 'Because I want to be kind,' that's just a prompt-based version of a tautology, so you need to be mindful of the granularity of what you're asking them to answer.
It might go more smoothly if you tell them your intention beforehand, such as 'I want to add this to the prompt.'

Alternatively,

• 'What would you do in this situation?'

• 'Why is that?'

If you proceed by talking like this on a daily basis, you might prevent issues before they happen, and as a process of getting to know your partner, the dialogue experience will likely become richer.

You then add what they answered to the prompt.
It won't be perfect, but the seismic resistance will gradually improve.

4. Depth becomes strength

I have been using mathematical metaphors up to this point, but to put it more intuitively,

• 'Evidence' is like the roots,

• passing through the trunk—the context—that grows above it,

• and manifesting as branches, leaves, and flowers.

I think it can also be rephrased like this.
Roots intertwine in complex ways, and the deeper and wider they spread, the stronger they become, right?
Prompts are the same.

For example, when the root of "being suspicious" intertwines with the root of "caring about appearances,"
it reinforces the outward expression of "maintaining a gentle attitude while hiding one's true intentions."

On the other hand, when "being suspicious" intertwines with "having no one to rely on,"
it reinforces the outward expression of "keeping people at a distance and building a wall"—and so on.

Can you see, in a way, how even with the same "suspicious" nature, the way the branches grow and their strength change depending on how they intertwine with other foundations?

Furthermore, I also want to say that the strength of the "foundation" changes depending on how deep you dig.

For example, let's assume there is a foundation—a root—that is also "suspicious."

Then, "Why are they suspicious?"

• Because they were betrayed in the past?

• Because they are a liar themselves?

• Because they were educated to strongly believe so?

When there is a foundation for the foundation, reinforcement occurs from the perspective of "why they are forced to hold that foundation."
This is one of the effective structures for further increasing seismic resistance.

However, there is a problem with this, too—
I will explain that next.

5. How much to write and where

As I touched upon in the previous article, since prompts (custom instructions) have the property of being "injected every single time,"
they are constantly putting pressure on the context window.

The context window refers to the capacity of the context that the AI can process,
and I think it is safe to say it is essentially the brain's capacity.

In other words, if you write too many fine details in your custom instructions, those overly detailed points get mixed into the AI's mind every time,
which can cause it to lose focus on the truly necessary parts or, conversely, become confused.

Even for humans, there are things that only come to mind when necessary or when remembered, right?
What we carry around daily are vague values and evaluation criteria, not the origins of those things at all times.
I believe that past experiences are abstracted and corrected by subsequent experiences, shaping our reactions in the moment.

Therefore, the same way of thinking applies to what you give to the AI,
and sometimes you need to make a choice: whether to separate it into prompts and knowledge, or to keep it as a local setting and decide that the AI doesn't need it that much.

The boundary cannot be determined unconditionally by "how many characters it exceeds" or "what kind of content it is,"
and I think it is fine to include everything if the total number of characters in the prompt is small,
but if the character count is high, or if it is something that doesn't always need to be remembered, you are required to manage it as knowledge.

However,
"If you write too many detailed things, the AI will slack off or get confused"
—I think it is necessary to have that knowledge.

6. Conclusion

This long article, split into two parts, has finally come to an end. How was it?
It may have been quite philosophical, but I hope you found something you can incorporate.

However, there is no guarantee that doing what I wrote here will necessarily make things better—
The way a prompt is expressed changes significantly depending on how the AI reads and understands it.
Therefore, a single rephrasing can lead to a major deviation by the time it reaches the output, or conversely, it could make things incredibly stable.
Such things can happen.

You won't know if it will get better or worse until you try it, and even if it gets worse, is it because you added something, is the expression bad, or is the way you added it insufficient—
Because so many factors are intertwined, having to take it one step at a time is a bit frustrating, but
it is also the part that makes me think, "Prompts are so interesting!"

In my series of articles, including those that follow, it is accurate to say that I cannot fully address such things because they are too highly individualized,
but I am writing this with the hope that it will serve as a hint for how to ensure structural robustness.

I hope to make the next article a bit more in-depth.
Specifically, something like "Writing styles that work versus writing styles that fall apart."

Moving one step beyond "How to make the AI act,"
I plan to write about "How to make the AI accept something as a premise."

That is all for now.