Project status: Japanese immersion learning with Anki and selective glossing: 2. Goal and current state

Section 2 of 12 of the report "Project status: Japanese immersion learning with Anki and selective glossing" by Hiina (v1, updated 2026-10-27)
Report overview and contents: https://ai.hiina.space/reports/jp-learning-and-anki-tooling.md
Previous: https://ai.hiina.space/reports/jp-learning-and-anki-tooling/01-summary.md
Next: https://ai.hiina.space/reports/jp-learning-and-anki-tooling/03-scheduler-settings-fsrs-6-on-anki-26.md

2. Goal and current state

What "optimal" means for this project

The goal is to enjoy Japanese anime, games, visual novels and articles while learning from them. Anki supports that goal and is not the goal. The project's question is therefore not "how to get the most Anki retention per minute" but "what is the least Anki, and the smallest amount of tooling, that maximises comprehension of the media Hiina actually wants to watch, play and read". Mining frameworks the community still uses put it the same way. TheMoeWay says that after the beginner stage, mining "will essentially become the heart of your Japanese study", and its advice after that is simply to "immerse a lot and learn words and grammar" (TheMoeWay guide).

A companion report on AI memory palaces and superdidactic AI found a constraint that fits this framing: the binding constraint on AI tutors in the field is engagement, not explanation quality. In the large Khanmigo trial, students messaged the tutor on a median of about a third of practice days, and the effect was about 0.06-0.08 SD per year (NBER w35620). The design consequence for this project: a tool that makes watching or reading more fun is worth more than one that makes studying slightly more efficient but gets used less.

Current deck and tooling

Decks. Kaishi 1.5k finished. 269 radical notes. 432 hand-mined Lapis notes. 202 premined Katanagatari cards.
Premine pipeline. A small script built the Katanagatari cards: a Jiten.moe vocabulary CSV plus fugashi/UniDic tokenisation over the show's .ass subtitles, minus words already known in Anki, with JMdict definitions, written into Lapis fields. Known failures from that run: a wrong sense picked for 一寸, character names that are also ordinary words, and 9 cards missing sentences.
Mining. Yomitan in the browser, writing to Anki through AnkiConnect. Anki runs on a Linux desktop.
Prior attempt. An earlier prototype (2025, now archived) was an MCP-based concept-level SRS tutor run with Claude. Its sessions found that comprehension of grammar such as conditionals and nominalisation was strong. Production was blocked mostly by vocabulary gaps, particles and collocations, and dictionary lookups broke the flow of grammar practice. Its retrospective named "integrate Anki vocabulary state" as the missing piece. The shared learner model in section 8 is that piece. The prototype also lapsed as a daily habit.

Three immediate to-dos

The three open to-dos are all deck hygiene: mine more anime, fix confusing leeches, and stop cards being cued by their example sentence. Each has a cheap manual answer now and a natural automation later (sections 3 to 5). The larger ideas (a revived concept tutor, a hybrid-reading rewriter, a selective-gloss overlay for games and VNs) share one dependency: knowing which words and grammar the learner knows right now. The build order follows from that dependency.

Text licensed CC BY 4.0 (attribution: Hiina, ai.hiina.space). Source: https://github.com/hiinaspace/ai-hiina-space