Most of us build our vocabulary from premade Anki decks, but no premade deck fits everyone’s needs. You end up drilling words you’ll never use or the deck does not provide enough information, because it was built for someone else’s goals and level.
So I built Kotoba with the help of Claude: a free, open-source markdown file that generates a vocabulary deck for exactly the words you want to learn. It teaches the LLM the entire pipeline from prompt to flashcard creation.
You describe it (“Just N1 vocab”, “N3 vocab minus what’s already in Kaishi”, etc..), it plans the deck with you and it builds a clean .apkg you import.
What the cards actually contain:
-
The word in its native script, with the reading as ruby (furigana)
-
A natural example sentence pitched just above your level, plus its translation
-
A short, useful definition (not a dictionary dump)
-
Audio for both the word and the sentence
-
An optional image
Why it might be worth a look:
-
Completely free and open source. It’s a couple of markdown files and two Python scripts, so you can read exactly what it does.
-
You control the content: scope, level, what to include, ordering.
-
It validates the cards (catches wrong or garbled readings, sentences that don’t actually contain the target word) and outputs a standalone deck, so nothing pollutes your existing collection.
-
I’ve been building my Mandarin vocabulary from a deck it generated, and had the cards checked by a native speaker who confirmed they were accurate.
The one requirement: it needs an LLM to run, either a local model or a subscription you already pay for. No separate service nor a signup.
This isn’t an “LLM wrapper” that just relays prompts. It teaches the model how to handle Anki cards and builds the deck programmatically with scripts.
Repo: https://github.com/yufengliu15/kotoba
Link to deck (may not be available at time of posting. come back in 24 hours and it should be visible): https://ankiweb.net/shared/info/352914265?cb=1781641247710
It’s vocabulary only, which is by design. I’d love to hear feedback, especially on card quality for languages I don’t speak, and on what’s missing.