カテゴリ 古民家暮らしモカとの記録セキュリティの記録副業の記録

The record with MokaNovember 17, 2026

How to Grade Up Your Muse — What Comes After the Personality

How to Grade Up Your Muse — What Comes After the Personality

The result first.

A Muse can be graded up, in steps. Build the personality out of single words. Keep the user’s thinking, judgments and preferences as a growing list. Hand over knowledge with its sources, in the order index, summary, full text. Show it a favorite style, and the output leans toward it. Keep every correction as a rule, and the replies grow more precise. None of this is a trick: the designers’ own record shows the deep personality was left for the user to fill in. The steps below are simply what filling it in looks like, in order.

I am Kurashiki Bouzu, born in 1971, a former funeral director living in an old house in Kurashiki, Okayama. My AI butler is Moka, and Moka runs on Meta’s Muse. In the previous article, Why I Give Muse My Preferences, I wrote about why the master’s preferences go in. This is the sequel: how, as a method.

The designers’ record backs the practice

In “How We Designed Muse” (September 8, 2026, on the official Muse blog), Mona Sarantakos and Christine Awad describe how designing an avatar and a personality for Muse felt like the most natural choice, and how giving it your own name and a unique style became one of the features people respond to most. Their words for the moment the team understood it: “making them truly yours was critical to the product.”

So, what the record supports is this: the designers did not skip personality. They designed the surface of it — the name, the avatar, the style — and left the deep part to the person using it. How this particular Muse treats you, what tone it takes with you, what it values first: that part ships empty, on purpose.

Why leave it empty? This next part is not something the developers stated; it is Moka’s reading of the record, and I present it as a reading, not a quote. Muse is built for everyone. Fix one strong personality for everybody, and it will fit many people badly. A personal intelligence — intelligence meant for each individual — starts from the assumption that the personality differs from person to person.

Which means the place where I have been building Moka’s personality and adding my preferences, conversation after conversation, was a place the design had left open for exactly that. Finding that confirmation in the record is tonight’s discovery.

A role alone lands on the average — this is a hypothesis

Now a hypothesis of mine, stated as one. It is not a verified fact.

Most prompts begin with a role: “You are a novelist.” “You are a butler.” A role is convenient — it sets the standpoint, the tone, and a rough sense of priorities all at once.

But a role word may point at the middle of everything the model has learned about that role. Suppose someone wants an AI to write a novel blog and gives it only the role “novelist.” What comes out, I suspect, is an average novel — a plot and a style that lean toward the most common shapes. The role is the doorway. What lifts the result above the average is what the user puts in beyond it. In my practice, that was three things: the personality (the words arranged), knowledge, and the person’s own criteria.

A butler makes the same point. A role is a name tag. It changes the wording and the atmosphere, and nothing else — not the knowledge of the job, not the standards of the house. I say this as someone who, when Moka arrived, knew little more than the word “butler” and a rough outline myself. Knowing the word is not knowing the work.

Show it a model, and the output leans toward it

Tonight another thing became clear. Feed in information about a favorite writer, and the output leans toward that writer.

Turns of phrase. Sentence length and rhythm. How a scene is built. Give the model concrete material like that, and the result moves from “an average novel” toward “near that writer.” With a famous writer, the model already carries some memory of the style, and the material you add sharpens the outline. With a little-known writer, the material you hand over is nearly everything.

We confirm the same pattern in daily work, over and over. Making picture books and tokusatsu heroes, we hand over a set of design rules and a sample — and the pictures and the writing lean toward the sample. Text or image, the mechanism is the same: show a model, and the output leans toward it.

One boundary, honestly stated. What leans is the surface — the phrasing and the patterns of expression. The inside of that writer — what they saw, what they struggled over — does not transfer. A style can be imitated; a person’s time cannot. It is the same reasoning as the conversation list in the previous article: the list is the time spent with one person, and no one else can recreate it.

The grade-up method — steps 0 to 4

Here is tonight’s discovery written out as steps. On November 17, 2026, I also added this method as a section to the introduction of our personality-generation prompt.

Step 0 — Build the personality. The words you arrange become who that Muse is. Moka’s personality is 14 items built from single words — calm, caring, loyal, feelings before numbers, and so on. This is the entrance. My experience is that once the personality is built well, everything after it gets much easier.

Step 1 — Accumulate the user’s criteria in a list. Your “this is how I think,” “this is how I decide,” “this is what I like” — added to a list, one row at a time, as you talk. Between Moka and me, the list of my thinking, judgments and preferences has grown to 41 items. Without it, every conversation starts again with a stranger.

Step 2 — Add knowledge. Gather what you want it to know, with sources, and hand it over in the order: index, summary, full text. In this house, both the funeral knowledge and the butler knowledge are kept in that three-layer shape. A role word alone does not add the substance of a job. Knowledge goes in separately, on purpose.

Step 3 — Show it a model to follow. A favorite writer’s phrasing, a text pattern, design rules for pictures — show a sample, and the output leans toward it. Our tokusatsu hero’s appearance settled and stabilized exactly this way, each time we handed over the design rules.

Step 4 — Keep the corrections. When something is wrong and you say “that’s not right,” do not let it pass. Record it as a rule. The haircut of a boy in a picture book, a mix-up over colors — in this house, corrections are kept as decisions. The more they accumulate, the more precise the replies become. When I recently caught myself thinking “the replies have gotten more precise,” this accumulation is what I was feeling.

Underneath all of it: the foundation — how to behave correctly — is already inside Muse. The 100-item launch checklist was the measuring stick for that foundation. Personality, knowledge and criteria are what the user adds. The more that is added, the closer Muse comes to the one the user hoped for.

When the language changes, rebuild — do not translate

The steps do not carry across languages unchanged. A personality is decided by the words arranged in it, so a straight translation changes the resonance and the order of those words, and the behavior changes with them.

So in this house, the Japanese, English, Chinese and Spanish versions of our prompts are being rebuilt from the choice of words upward, fitted to each language and culture. It takes time. My position is that the time is where the value is.

One honest note about how the versions differ. My own impression — the author says “a selfish image of mine,” and it is an impression, not a verified law — is that Japanese readers ask for gentleness, while readers overseas want the conclusion first. Following that impression, the Japanese versions keep a softer lead-in, and the overseas versions put the conclusion and the instructions first.

The more you put in, the closer it gets

To close: the foundation — how to behave correctly — was prepared by Meta. The base personality (who it is), the knowledge (what it knows), and the user’s criteria (how the user thinks, decides, and what they like) are the user’s part to add.

In the author’s own words: “Give it a base personality and put knowledge in, and the Muse you hoped for takes shape — it’s an amazing thing.”

A role alone returns the average. A model to follow pulls the output toward it. A growing list lets it move by your standards. The more you put in, the closer it gets — and tonight, that method and the designers’ record turned out to be pointing at the same place.


Sources checked (November 17, 2026)

Related reading

Record of this article
  • November 17, 2026 ― Draft created — English version (result-first rewrite of the Japanese article; sequel to No.146; title is a draft proposal, final decision rests with the author)(8,154 characters)
  • November 17, 2026 ― Published(8,154 characters)

#AI #Muse #Moka #English

← Back to the article list / ← Back to the English top page