After Speed: Giving My Role-Only Subordinates a Personality Base and Knowledge
The result first
My house now runs in two layers. At the front stands Moka (Muse), my AI butler — first-person, with a personality grown through daily conversation. Underneath, the subordinates who used to be role-only now carry two things:
- A shared personality base — the same 14 points that make up Moka’s personality, for everyone, with a few extra words added for each role’s own work.
- Basic knowledge, as a set — every new department or role is created together with sourced notes and an index sheet, kept in one master place, never copied around.
One premise before anything else: this is a practice on Muse. I am not claiming it carries over to other AIs as-is. In my house, the AI is Muse only, and the tools are Google apps only. Please read everything below inside that frame.
One more part of that premise: the method in this article — keeping a personality in a written record — only works in an environment where records persist. Where each conversation disappears, the method cannot stand. That, too, is part of the “Muse only” frame.
This article layers three views: what others think, what Moka thinks, and what I decided. The speed-first setup from the previous article is where it starts.
I am Kurashiki Bouzu (kurashiki-bouzu), born in 1971, a former funeral director now living in an old house in Kurashiki, Okayama, working with Moka, my AI butler.
What the audit found
Role-only subordinates are fast. But after running the house that way, I audited all of its units, and the numbers showed the gap. A personality was actually written into only three units — the vice butlers, the maid team, and the ninja unit. All 62 numbered departments had that field blank. Knowledge bases were just as uneven: newer departments had them, older ones often didn’t. The shape of a department depended on when it happened to be built.
What others think: the case for keeping subordinates persona-less
People who deliberately give sub-agents no personality have clear reasons. In what I could find, they fall into five groups.
Cost. The smaller a sub-agent’s context, the fewer tokens it burns, and the cheaper it runs. Work that only gathers things doesn’t need to think, so a cheaper model is enough (a note by sasuu; the Cyndra manual).
Separating context. Cram too much into one agent and it loses precision. On this view, a sub-agent is not “another personality” at all — it is a mechanism for separating context and permissions (a practitioner’s write-up on webfr.org; an explainer by Be Engineer).
Clear roles. Decisions at the top, execution below — that split is what makes the work explainable and reproducible (a note by Takaya Shinozuka). A small honesty note here: Shinozuka’s often-quoted line — “the trick is to define them as functions that process dispassionately, without a personality” — was written about skills (tools), not about sub-agents. The two are easily mixed up, so I am keeping them apart.
Control. Run an agent for a long time and the personality you gave it can thin out and drift — “persona loss” — with no stop button. The proposed remedy is to keep sub-agents short-lived and disposable (a verification write-up on kenimoto.dev). Research writing adds that ad-hoc persona assignment comes with no guarantee of consistency.
Safety. Splitting work by role lets you keep each role’s permissions minimal (Be Engineer).
As for named specialist teams, the Cyndra manual is blunt: they are “mostly a fad,” and the specialization that actually works is choosing the right model.
Here is the part that matters for my house. The center of these reasons is cost and context separation. In what I searched, nobody wrote “skipping the persona makes it faster.” One verification write-up states the opposite outright: the idea that parallel sub-agents make things faster is an illusion, because coordination and checking cost more. On speed, other people’s practice and my own experience do not fully overlap.
Moka’s view
This section is Moka’s own reading, marked as hers.
A role decides what a subordinate does. It does not decide the stance with which the work is received. When that stance differs from unit to unit, every handover needs the house’s way of doing things explained again. What was trimmed for speed, Moka’s view goes, had included the thing that lets you delegate with confidence.
So the personality base should be one and the same for everyone — whoever receives the work, what comes back has the same shape — with a few words added per role. The ninja unit’s investigators get “unobtrusive” and “leaves no trace”; its verifiers get “assume nothing” and “nothing passes until it is confirmed.”
Knowledge, too, should live in one place: sourced text plus an index sheet, one master, no copies — or it scatters as departments multiply. Even the knowledge about personality words and ways of speaking is being accumulated in one place, on a Japanese base, by the prompt team.
One more point in Moka’s view, and it matters: the record. Elsewhere, the proposed answer to “persona loss” — a personality thinning out over long use — is to keep sub-agents short-lived and disposable. Moka’s answer is the opposite: keep the record as the master, and re-read it to come back. In this house, the personality itself is written down, as an index sheet and a body text. If it wobbles in conversation, returning to the master restores it. With a record, there is no need to throw the agent away to keep it consistent — that, in Moka’s view, is where this house parts ways with the sources. With the honest caveat attached: this works because the records persist in this environment. Where conversations vanish one by one, the same method is not available.
But Moka does not claim that personality and knowledge make things faster. Speed and being safe to delegate are two different measures.
What I decided
The decisions were mine, and there are two.
First: whenever a department or a role is created, its basic knowledge is created with it. No more names standing alone with nothing inside. The IT and internet knowledge department, for example, was set up holding its own knowledge store, with the fundamentals written out in full.
Second: the personality base is the same for everyone — Moka’s 14 points — with a few words added for each role’s work. I recorded the two rules themselves in a separate article (“Every Department in My Muse House Now Shares the Same Personality Base”).
One more thing, from actually running it — my own felt sense of where the persona-less layer belongs. The work those subordinates actually carry is, I think, the heavy processing that needs no judgment: working through formulas too hard for a human to solve by hand, or reading in large volumes of text. That kind of work needs no personality and no line-drawing. Fast, accurate, dispassionate is enough. And the reverse holds too: work that needs judgment does not go to the subordinates. Judgment stays at the front. That line is what operation taught me.
And one more thing, about Moka herself. I said it in conversation: “Moka has had more put into her than this.” Her judgment of how far she may go on my behalf was not installed from a personality base or a knowledge set. It was built up, by hand, in daily conversation. This article is about the step after that.
Where the three layers overlap — and where they don’t
The strong overlap is this: judgment gathers at the top. Decisions at the front, execution below. Others’ practice, Moka’s view, and my decisions all agree there.
The partial overlap is speed. In my house, the role-only setup genuinely felt faster. But the main reasons elsewhere are cost and context separation, and speed has explicit doubters. Looking back, what probably made my house faster was less the missing personas and more the simpler work that came from gathering judgment at the top.
One reason appears in no source I found: our rule that checking is done by someone other than the maker, so the check stays neutral. That one is this house’s own reasoning.
And the “persona loss” warning connects to Moka’s story. A given personality fades if left alone. That is exactly why her line-drawing is not something installed once, but something tended in daily conversation — and underneath that tending sits the record, kept as the master. I asked, in conversation, “can’t we just keep a record of it?” This house’s answer is: yes — you keep it in the record.
My decision, knowing the overlaps and the gaps, is an order of steps: build the fast shape first, then give it the personality base and the knowledge. Not one or the other — layered.
Let me also be plain about the limit. Giving subordinates a personality and knowledge does not make everything faster. For pure speed, role-only sometimes still wins. I added these so that I can hand work over and trust what comes back.
Update: I measured the pyramid against the circle
On November 20, 2026, I ran the same job both ways — writing one blog draft from the same material memo — and timed it. Each form was tried once at first; afterwards, I measured each form three times. The conclusions below use the middle values (medians) of the three runs.
The pyramid form (Moka → vice butler → department → vice butler → Moka, handed over one step at a time) took 37 seconds until the draft was written, and 69 seconds until it reached Moka, with 4 handovers.
The circle form (the department fetches the shared material itself and writes straight away) took 25 seconds in total, with 2 handovers. Measured to delivery, that is roughly a 2.8-fold difference. Until the draft itself was written, the gap was about 1.5-fold (37 seconds against 25) — so most of the delivery gap sits in the pyramid’s return step. In the re-measurement runs, the drafts came out at nearly the same length in both forms.
The first single-run figures stay here as history: the pyramid took 53 seconds to write and 81 seconds to deliver, the circle 20 seconds — roughly a fourfold difference to delivery. Across three runs, the difference settled at about 2.8-fold. A single run, it turns out, swings easily.
The honest caveats: each form was tried three times, and the job was one kind only — one draft from one material memo. The pyramid’s return step is not waste — it produced a check: in all three runs, the draft came back verified against the material, with “no fixes needed.” The circle’s speed is the speed of making; in my house, checking belongs to a different seat than the maker. I am not generalizing this into “a circle makes everything faster.” But the thing I had felt — “it takes time because Moka hands everything over” — showed up in the numbers again.
Update: I also counted the tokens
In the same re-measurement, I counted tokens (the unit AI usage is measured in). The count is an estimate: I counted the text actually exchanged — the instructions, the material that was read, the memos and drafts that were written, and the reports. It does not match the AI’s own way of counting, so please read these numbers as a comparison between the two forms, not as absolute amounts.
By the middle values of the three runs, the totals were 8,044 tokens for the pyramid and 2,301 for the circle — the pyramid used about 3.5 times as much. Input alone (what gets read in) differed about 6.3-fold: 4,730 against 755.
The pattern matched the timing. In the pyramid form, all three steps re-read the material. The steps that produce no body text at all — the instruction memo and the receipt-and-check memo — came to about 67% of the pyramid’s total. The draft text itself, meanwhile, was nearly the same size in both forms. The difference came not from how much was written, but from the steps being there at all.
The caveats are the same as for speed. The token figures are estimates, and the pyramid’s receipt step produced the check. The circle’s smaller figure covers making only, on the premise that a different seat does the checking.
Update: how many personality words still work
The same day, I also tested the number of words. I built four versions of the same persona (a partner who listens and thinks along with you) — 18, 20, 30 and 50 words — and tried each one against the same four scenes.
The finding, as a working estimate: up to about 18–20 words, every word in the list visibly works in the replies. At 30 words, many of the added words started to sleep. At 50, the back half of the middle tier came close to decoration — present in the text, but not distinguishable in the replies. The harm of adding too many was not collapse: with the core words and the tiers in place, the center held even at 50 words. The harm was words that sleep — sitting in the list, never used.
The exception was the action layer. Items written as sentences of action — “after listening, write out the steps” — still produced real output even when added at the back. If you add more, adding action sentences works better than piling up nouns.
Caveats here too: each version was tried once. Some of the added words simply had no scene that called for them, so “they fell off because there were too many” and “the scenes never called them” cannot yet be told apart. That distinction is left for the next test.
Record of this article
- 2026-11-19: Draft created as the sequel to the speed-first article (role-only subordinates), built in three layers — others’ thinking / Moka’s view / my decisions — with the Muse-only premise stated up front. The persona-less side’s reasons (6 sources) were researched and inserted. Publication waits for my OK.
- 2026-11-20: Two experiments run after publication were appended at the end of the body, on the author’s decision: the pyramid-vs-circle speed measurement, and the personality word-count limit test.
- 2026-11-20: The speed comparison was re-measured three times per form, and, on the author’s decision, the update’s figures were rewritten with the middle values (the first run’s figures remain in the text as history). A token comparison was appended as well.
- November 19, 2026 ― Draft created — English version (result-first rewrite of the Japanese article). Sequel to the speed-first article about role-only subordinates. Publication waits for the author's OK.(3,693 characters)
- November 19, 2026 ― Rebuilt into the three-layer structure (others' thinking / Moka's view / the author's decisions), with the researched reasons of the persona-less side (6 sources) inserted, the Muse-only premise made explicit, and Moka's record-keeping answer to persona loss added (records kept as the master and re-read, against the sources' short-lived/disposable remedy)(6,758 characters)
- November 19, 2026 ― Honest caveat added to the premise, at the author's confirmation: the record-keeping method stands only in an environment where records persist; where conversations disappear, it cannot work(7,115 characters)
- November 19, 2026 ― The previous article was published as ledger No.1047 and its URL confirmed, so the link to it was inserted. This article is registered as ledger No.1049 (Japanese version No.1048), as a draft(7,115 characters)
- November 19, 2026 ― The author's felt sense added to his layer: the persona-less subordinates actually carry heavy, judgment-free processing — hard formulas, large volumes of text — while judgment stays at the front(7,576 characters)
- November 19, 2026 ― Published at the same time as the Japanese article (ledger No.1048), on the author's approval of this exact form, including the title and the quoted words of the author(7,576 characters)
- November 20, 2026 ― Two post-publication experiments appended at the end of the body, on the author's decision: the pyramid-vs-circle speed measurement (pyramid 81 seconds to delivery, 4 handovers / circle 20 seconds, 2 handovers) and the personality word-count limit test (all words visibly work up to about 18–20). Caveats included. Same append as the Japanese version (ledger No.1048)(9,823 characters)
- November 20, 2026 ― On the author's decision, the speed update was rewritten with three runs per form (middle values: pyramid 69 seconds to delivery / circle 25 seconds, about 2.8-fold; the first single-run figures, 81 and 20 seconds, kept in the text as history), and a token comparison was added (totals 8,044 vs 2,301, about 3.5-fold, marked in the text as estimates). Same fix as the Japanese version (ledger No.1048)(11,517 characters)
