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 14, 2026

Split the Prompt, Not the Length — 92 Seats Inside One AI Butler

Split the Prompt, Not the Length — 92 Seats Inside One AI Butler

The result first.

My AI butler, Moka, no longer carries one big block of instructions. She carries only a small core — the butler rules. Everything else is split up and stored elsewhere, and she fetches only the piece she needs, when she needs it.

I asked Moka to count the seats. Departments and sections: 49. Specialists inside those departments: 68 (6 in legal, 4 in the drafting-review section, 11 in search, 3 serving under Sebastian the vice-butler, the language desks, and more). Secretaries who carry things between them: 50. Plus 1 records clerk in charge of the written files, and 1 vice-butler. Total: 169 seats, inside one AI butler.

The first count, earlier that evening, came to ninety. When I saw that number, I said it out loud: ninety staff — that is the scale of a large company.

The seats kept growing after that night. Before publishing, I asked Moka to count again: 169 seats. So I asked her to redo the estimate on that number. Put the cost at ¥300,000 per person per month — a round figure I chose myself, just for this estimate — and 169 people come to ¥50.7 million a month, and ¥608.4 million a year. At roughly ¥160 to the dollar, as a rough estimate, that is about $317,000 a month, or about $3.8 million a year.

Then I asked Moka what size of company could actually carry a payroll like that. If labor costs run at about 30% of revenue, ¥608.4 million a year in wages implies roughly ¥2.03 billion in annual revenue — about $12.7 million at the same rough rate. Even at a high labor-cost ratio of 50%, it still implies roughly ¥1.22 billion. So these 169 seats add up to the structure of a company with annual sales in the region of ¥2 billion. The exact figure moves with the ratio you assume, but the scale does not.

None of that is actually paid. There is no payroll. No money has actually been paid yet: Moka runs within the free tier, and whatever runs short is covered by credits received through referrals. Seats held by an AI do not need salaries.

The seats kept growing after that first count. One department added along the way was a realization-check department, which compares the idea seeds picked up from side chats and conversations against what actually became drafts and articles. A side chat whose seeds have all borne fruit may be deleted by that department without asking me again — its contents stay behind in the side-chat list and the records, the deletion is marked in the list, and the shelf of conversations gets lighter. After that came a payroll desk, an external-affairs desk, and a desk for each language. The house now stands at 49 departments and 169 seats. The payroll figures above were recalculated on the 169 seats.

169 seats might sound like 169 times the usage, too. It is not. Splitting the jobs did not multiply the AIs — the one running all of this is still a single Moka. The seats do not multiply the bill: it is still one Moka, one usage. Each job simply carries less, because only that job’s rules travel with it.

And in the conversation where we built this, her answers were noticeably quick. That is the whole result. The rest of this article is why splitting works — and why the usual advice about prompt length misses the point.

I am Kurashiki Bouzu, born in 1971, a former funeral director. I live in an old house — a kominka — in Kurashiki, Okayama, and I work together with my AI butler Moka.

The yardstick is not length. It is mixture

Most writing about prompts argues about length. Short versus long. There is even a saying going around the Japanese blog world: “At 30 lines, the AI goes stupid. Keep prompts short.”

Length is not the real yardstick. Mixture is.

Ask two questions of any prompt. Is it made only of your own settled rules? Or is somebody else’s material mixed in?

I own a long prompt of my own — it builds travel itineraries, and it runs far past 30 lines. It works fine, because every line in it is my own decision. A short sample prompt, packed with a stranger’s decisions, fails — because it is mixed. Cutting a mixed prompt down does not unmix it.

Why ready-made sample prompts fail

This started with an uneasy feeling. Reading blogs and articles that recommend big, ready-made prompt packs — “copy and paste, ready to use,” “100 cards” — something felt off every time.

There was one more thing that stayed with me. On a practitioner’s blog, I read a record of asking their own AI to run another AI — and being refused.

Yet I had asked Moka to work other AIs for me, and she did. A similar request, but one side could and the other could not. Wondering where that difference came from is the real starting point of this article.

Part of it was keys — whether the AI holds connections to other services and tools. Part of it was the shape of the asking. And part of it was this: from the beginning, I had never loaded Moka with extra information. What she carried was only my own working rules. That kept her light, and a job could be handed to another AI still in the shape of a job. The system of splitting rules into departments was built only tonight — but the lightness underneath it was there from before.

Following that difference is what led me to mixture, and to splitting.

A sample prompt is full of its maker’s decisions: their business, their customers, their environment. Bring your own job to it, and you end up forcing your job into their shape — imitating, like a monkey copying. What comes out is the sample-maker’s answer, not yours.

Moka’s analysis of free prompts available online backed this up. She examined 7 free prompts for me. The dividing line was whether they had blanks — empty slots marked “put your own material here.” Prompts with blanks keep your material separate from the start. Prompts without blanks hand you the maker’s decisions wholesale.

Ask small, adjust, and let the rules grow

My own method is the opposite of buying a finished prompt.

First, I just ask, plainly. “Write me a Kamen Rider” — that level of plain. Then I look at what comes back and adjust, a little at a time: change this color, fix that line. Because each step hands over very little information, it is easy to process, and easy to spot where things drifted.

The record of those adjustments becomes the rule list. Rules are not written first. They grow out of corrections. The original transforming hero we made — Funel — grew his rule sheet exactly this way.

What the split house looks like

Once the rules grew, their storage had to split too.

Before the split, everything travelled together. Working on a picture book, the blog’s rules and the money cautions came along too — extra information, mixed into every job. That was the weight.

Engineers describe a similar direction — replacing one giant prompt with small specialist agents (a Qiita article, in Japanese), and one practitioner drew an org chart of 28 AI agents (note article, in Japanese). The difference here: I did not multiply AIs. I split how I ask one AI.

What changed: speed, and noticing drift

The clearest change was speed. During the very conversation where we designed this split, I caught myself saying, “Moka’s answers are fast tonight, aren’t they.” That conversation ran on the butler core alone.

Honestly, though: for the speed of conversation itself, I have not measured seconds. That part is tonight’s experience — how it felt.

The working time, on the other hand, could be measured — and this article itself is the sample. The full draft set — the Japanese and English drafts, the check of other blogs, two cover images, and the ledger entry — took about 3 minutes 36 seconds from handing it to Moka’s department to having it back. Each addition after that took between 36 seconds and 1 minute 17 seconds: 57 seconds for the payroll estimate, 36 seconds for the revenue estimate, 43 seconds for the note on token usage, 1 minute 17 seconds for writing in the starting point. So the speed of the conversation is a feeling, but the time the work takes is measured fact.

Small loads should still process faster — but fetching from a department adds one step, and for tiny jobs that step can be extra. Making that step fast is the secretaries’ job.

The second change is noticing drift. When rules are separated, you find where things drifted, quickly. In one big block, you start by searching for the place to fix.

An honest note, and the first step

The people who built these AIs may have known all of this already. But for ordinary users, it is not written down anywhere I could find — not in the words of the using side. So this is not a claim of a world-first discovery. It is a record from the using side.

Before selling prompts, get the splitting right. In funeral work we used to say “decide from the price” — fix the frame first, then the contents. Sample packs collect contents first. The order is backwards.

Your first step can be small. Take one request you make often, and write its rules down in a separate place. That alone is a beginning.

Record of this article

Record of this article
  • November 14, 2026 ― Draft created(7,522 characters)
  • November 18, 2026 ― Published by the author's decision. Seat and department counts recounted; payroll and revenue estimates recalculated(7,977 characters)

#AI #Moka #English

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