Features · 機能
Yuki isn't a wrapper with a prompt. It's six coordinated systems built specifically for questions about Japan — the register you speak in, the memory of what you're working toward, the shape of the answer you actually need.
01
Trust boundary
Scope guard
Every query is checked against a hard Japan boundary before it hits the model. Off-topic asks are politely declined — your credits and latency never leak on tangents.
02
Continuity
Per-user memory
Yuki remembers your target city, JLPT level, timeline, and dietary needs across sessions — pulled from a private, RLS-scoped table injected into every prompt.
03
Keigo / 敬語
Honorific register
Toggle between casual, desu-masu, and full keigo. Yuki adjusts particles, verb endings, and vocabulary — the same idea, said the way a Japanese recruiter reads it.
04
Not paragraphs
Structured responses
Roadmaps render as timelines. Itineraries render as day cards. Grammar renders as side-by-side examples. The model returns typed JSON; the UI renders it as first-class objects.
05
Router
Intent classifier
Every message is routed — visa, language, travel, career, culture — and shaped by an intent-specific system prompt so answers land in the right register and depth.
06
Latency
Token optimiser
Kanji is ~3× denser than English; Yuki exploits that and compresses history aggressively. First token in under 400ms on a warm session.
Six domains · 六道
005 · Contrast
“How do I move to Japan?”
Generic AI
Generic bullet list about visa types.
Yuki
A year-by-year roadmap keyed to your target date, JLPT level, and profession.
“Best time to visit Kyoto?”
Generic AI
“Spring or autumn.”
Yuki
Momiji peaks Arashiyama Nov 18–24 this year — book Higashiyama ryokans by mid-August.
“How do I write my rirekisho?”
Generic AI
Translated Western resume.
Yuki
Correct 履歴書 layout, photo spec, family register field, and keigo cover email.
Numbers
47
prefectures indexed
3
honorific registers
< 400ms
first token
100%
RLS coverage