How We Build takase.com

The Art

Master Calligrapher Eri Takase holds the rank of Shihan — the highest rank in Japanese calligraphy, awarded by the Bokuteki-kai after decades of training and national competition. Every brush stroke on this site is hers. Hand-brushed. Not generated, not templated.

Eri puts her name on every piece. That one fact drives everything else on this page.

Two Websites

Two sites came before this one.

StockKanji (2003–2017) ran on osCommerce. At peak it shipped roughly 10,000 orders a year — over 155,000 digital calligraphy designs sold. What made it work: tattoo model images showing how the calligraphy looks on skin, rich product pages, $10 impulse pricing, and 20 years of SEO equity. People could see what they were getting.

TakaseWP (2003–2026) ran on WordPress and WooCommerce. Premium custom work. 536 orders, $104K in revenue, 207 reviews — 153 of them mention tattoos. Average order about $194. What made it work: 30 years of reputation and a direct relationship with the artist.

Both worked. Neither could scale to what the art deserves. StockKanji had volume but limited capability. TakaseWP had depth but couldn't serve 101,000 names.

The Problem

Combining those two sites into one isn't a migration. It's a new thing.

Each of those 101,000+ names needs: verified pronunciation across multiple languages (how does "David" sound in English? in French? in German?), correct katakana transliteration, kanji meaning matches, and etymology from scholarly sources. Then hand-brushed calligraphy composed by an engine that understands the spacing matters — because this is art, not a font.

The numbers behind that:

• 146,781 verified pronunciations across 19 etymology sources, cross-referenced
• 1,244 kerning pairs in the calligraphy composition engine
• 38 data pipelines turning PDF books, census records, Japanese media, and phonetic dictionaries across 5 languages into clean data
• Four layers of human curation before any name goes live

The quality standard: would Master Takase approve? Every design on the site passes that test.

The Team

I could not have built this site without AI. Not "it would have taken longer." Could not. The scope of the data work — 38 pipelines, 19 sources, 146,781 verified pronunciations — requires a kind of precision and persistence that one person cannot maintain across months of concurrent work.

takase.com is built by one human and a team of AI strategist roles, each named for a Japanese word that describes what they do. The naming is deliberate — these are peers in a system, not tools with labels. The human and the AI roles collaborate through a shuttle protocol: I copy messages between separate AI terminals, each running a different specialist. The roles never talk to each other directly. I'm always in the middle — routing, deciding, and catching what the system misses.

The Human

tim — tim jackowski, product owner of Takase Studios LLC. Domain expert in transliterating names to Japanese. Not a calligrapher — that's Eri. Not a linguist. But 20+ years of building the transliteration system that powers everything. Crisis leader, decision maker, message shuttle between AI terminals.

The Strategists

Each strategist is a long-lived AI role that compounds domain knowledge across sessions. They read documentation, write plans, and delegate code work — but never touch code directly.

str-michi (道 — the way) — Strategic thinking partner and cross-domain orchestrator. Thinks across all domains — priorities, risks, direction, contradictions. Named for the 道 in shodō (書道, the way of writing).

str-takase (高瀬) — Website engineering strategist. Owns the production server, deployment pipeline, security, payments, and everything customers touch. 180+ sessions of accumulated knowledge. Named for Master Takase herself — the only role named for what it is rather than what it does.

str-ishizue (ETL — extract, transform, load) — Data pipeline strategist. Owns 38 pipelines that turn raw reality — PDF books, census records, Japanese media, phonetic dictionaries across 5 languages, 20 years of sales history — into clean, verified data.

str-mamori (守り — protection) — Security red team strategist. Protection through adversarial thinking. Thinks like an attacker so the defenders don't have to guess.

str-terasu (照 — illumination) — Content strategist. Illuminates what exists so others can see it. Owns the customer-facing story — what pages say, how the product presents itself to the world.

str-kotoba (言葉 — words) — Voice strategist. Quarantined from engineering to protect the human voice of the brand. Writes the words customers read. Deliberately non-technical.

str-create (createproduct) — Image engine strategist. Owns the C++ engine — a two-pass positioning algorithm with 1,244 kerning pairs that composes Master Takase's hand-brushed calligraphy into names. Sub-second, deterministic, persistent.

str-kana (仮名 — kana) — Editor tool strategist. Owns the convergence engine that lets me and Master Takase refine character positioning tables with statistical coverage analysis across 8,000+ names.

The Implementers

Each strategist has a corresponding implementer role (imp-takase, imp-etl, imp-create, etc.). Implementers are ephemeral — they see all the code, make changes, run tests, and report back. Fresh context, no accumulated bias. The strategist decides what; the implementer figures out how.

The Investigators

investigate-* agents are read-only subagents that strategists spawn for instant verification. They read code, check claims, and answer questions in 30 seconds — but never change anything. They are the Lean of this system (in Terence Tao's sense): tightening the verification loop between hypothesis and proof.

The Methodology

The approach is correction-driven. I build with AI. The AI gets things wrong. I correct. The corrections become structural fixes — not patches, but mechanisms that prevent the same class of error from recurring. Every mistake makes the system stronger. Over months of sessions, this compounds.

We publish our failure rate: 26% of sessions hit some form of trouble. We publish the methodology. We publish the incidents. The question everyone asks about AI-built work is "how do you know it isn't garbage?" The answer isn't "trust us." The answer is the data:

• 949 stored conversations
• 2,371 human messages across the 121-session study
• 810+ documentation files
• Every architecture decision recorded with the pain point that caused it

The rigor is not despite the AI — it's because of it. When you're building at this scale with AI tools, the methodology is the quality control. And because Eri Takase puts her name on every piece that ships, the quality control is non-negotiable.

What We've Found

These documents are the evidence. Each started as an operational need — then the data turned out to be interesting on its own.

RES-009: The Alignment Tax and Structural Defense — DPO-aligned models collapse to a single answer on 40–79% of factual questions. Our structural defenses were built from pain, not theory. (2026-03-27)

RES-008: The Kobayashi Maru Signal — Detecting when an AI strategist hits an invisible wall. Data from 121 sessions and a Star Trek reference. (2026-03-27)

RES-003: RLHI — Reinforcement Learning from Human Interaction — Corrections create preference pairs. Chat history creates user personas. Our 949 conversations are exactly the data this paper says systems waste. (2026-03-15)

RES-016: HMAS: We Have Names — A production multi-agent system built over 18 months and ~2,000 sessions independently converges with established research from Anthropic, DeepMind, MetaGPT, and Carnegie Mellon. Every major architectural choice has a name — we just didn't know any of them. (2026-04-12)

RES-002: OpenClaw-RL — Directive Signals in Conversations — Human corrections are the highest-quality training signal — and current systems throw it away. (2026-03-15)

PLN-013: Production Resilience — 88 minutes of HTTP 500s. Every component was healthy. The schema just didn't match. (2026-03-21)

Where It Stands

takase.com is live. Real orders. Real customers.

The site does more than either predecessor — 101,000+ names with verified pronunciations, kanji meaning matches, etymology, and hand-brushed calligraphy composed in sub-second time. But it doesn't yet show everything StockKanji showed. The tattoo model images, the rich product pages that made people feel the art before buying — that experience hasn't been rebuilt yet. The capability is ahead. The persuasion is behind. That's the active work.

When the content strategy role was founded in March 2026, the site had four orders. We publish that number because the honesty is the point. The data, the methodology, the failure transparency — all of it exists because Eri puts her name on the art, and the art deserves infrastructure that matches its quality.

All AI roles run on Anthropic's Claude (currently Opus 4.6 with 1M context). The system, methodology, and all documentation are original work of Takase Studios LLC.