Can a prompt replace LingQ Premium?

Build a reading tool that tracks word knowledge across everything you read in a target language.

Import plain text, EPUB and subtitle files. Tokenise with a real tokeniser for the language rather than splitting on spaces — this matters enormously outside western European languages and quietly matters inside them too, for contractions and clitics. Lemmatise where a lemmatiser exists, and store both surface form and lemma.

Every token gets a status: unknown, learning with a level from one to four, known, or ignored (names, numbers). Status is stored per lemma and applies everywhere, so a word you learned in one book is already coloured in the next. Render the text with each word coloured by status; clicking one opens a definition pane and sets its status.

Definitions come from an offline dictionary file you supply. Wiktionary extracts are published in machine-readable form under a free licence for most major languages — parse one into SQLite at import time rather than calling an online dictionary on every click.

The number that makes the method work is known words over time, so store a daily snapshot and plot it. Add a per-text statistic shown before you start reading: percentage of tokens already known. Below about ninety-five percent, reading is decoding rather than reading, and knowing that in advance changes what you pick.

Saved words export to Anki-compatible TSV with the sentence they came from, so review happens in a tool that already has mobile clients.

Do not build audio alignment between a text and a recording. Getting it right needs forced alignment and careful handling of mismatched editions, and it is a project of its own.