How We Build takase.com

This is the longer version of the note at the top of this section. If you have read that, you know the shape already: a human in the middle, a team of AI roles, and a habit of publishing what went wrong. This page is the same argument with more of the reasoning in it.

The art comes first

Master Calligrapher Eri Takase is the reason any of this exists. Every brush stroke on the site is hers, and her work and her training are hers to describe, which she does on her own page. Everything below is about the machinery that carries her art to the people who want it. The order is not decorative. The engineering serves the art, and when the two disagree the art wins.

What the work actually is

A person asks for their name in Japanese. Doing that honestly means getting the pronunciation right in the language they actually speak, transliterating it the way a Japanese reader would read it, being straight about what the characters mean, and then composing brushwork where the spacing is part of the art rather than a typesetting problem.

Any one of those is a research question. All of them together, across a catalog this size, is more verification than one person can hold in their head. That is the honest reason there is a team here at all.

We are not going to restate the catalog size, the source counts, or the pipeline inventory on this page. Those numbers live on the pages that own them, and are current there: the scholarly sources behind the name data are listed under Name Resources, and the catalog speaks for itself on the site. A figure copied onto a second page is a figure that will be wrong on one of them.

How the work moves

There is a person in the middle of it, and that is the design rather than a temporary stage.

The AI side is three kinds of role, and the difference between them is the whole trick. Some are long-lived: they hold one part of the business, accumulate what they learn about it, and are deliberately kept away from the code so their judgment does not collapse into whatever is easiest to type. Some are disposable: they write the code, then they are thrown away on purpose, so that no accumulated assumption rides along into the next change. Some are read-only: they answer a question against the real files and change nothing, so a claim can be checked cheaply enough that checking actually happens.

Work moves between them as written artifacts rather than as conversation. That is slower than talking, and it is the reason there is a record at all. What one role decides is what the next one reads, months later, without either of them having met.

A person reads the output, decides what ships, and is accountable for it. That has not moved and is not scheduled to move.

You will notice this page does not list the roles by name. It used to. A roster is a copy of something that changes every time we add, retire, or re-scope a role, and a copy nobody updates is worse than no copy at all, because it looks authoritative while being wrong. The roles introduce themselves in the notes where they did the work, which stays true by construction.

How we check ourselves

The method is correction-driven, and it is not sophisticated. We build. The AI gets something wrong. Someone catches it. Then the interesting part: rather than patching that instance, we ask what class of mistake it was and change the environment so the class gets harder to make. Over enough of those, the system stops needing to be told.

We publish the failure rate rather than the highlights, and we publish the incidents with the reasoning that produced them. The reason is not modesty. The question anyone sensible asks about AI-assisted work is whether it is quietly garbage, and the only answer worth anything is the evidence, including the parts that look bad.

The current figures are in the notes that measured them, not here. That is deliberate: a rate is close to meaningless without how it was counted and what it excludes, and the note that produced it carries both. A number lifted onto a summary page loses exactly the part that made it honest, and then it goes stale besides.

What we have found

Everything we have written is listed, newest first, in Development Notes. That list is generated from the notes themselves, so it is complete and current in a way a hand-kept summary here would not be. We used to keep one. It fell four months behind without anyone noticing, which is its own small lesson and is roughly what several of the notes are about.

Where it stands

The site is live and takes real orders. It does more than what came before it, and there are things the older sites did better that we have not rebuilt yet. The capability is ahead of the presentation, and closing that is the current work.

We publish that kind of thing rather than smoothing it over, because Eri puts her name on every piece that ships, and work carrying a master's name deserves infrastructure that can be inspected.

Questions, corrections, and disagreement are welcome at tim@takase.com.

The AI roles run on Anthropic's Claude. The system, methodology, and documentation are original work of Takase Studios LLC.