The tutor
Nine things a model is good at here.
And one rule that governs all of them: every one also answers with no model connected, from the seven hundred written grammar rules and your own review history. A language course that stops working when an API key expires is not a language course.
What it does
Nine jobs, in the order they matter
Why that was wrong
Being told the answer teaches the answer. Being told which rule you broke teaches the rule, and the rule covers a hundred more sentences. This is the single most valuable thing here and the thing that does not scale with human teachers.
Ask about the language
A question box in your own words. Why this form and not that one, when a region says it differently, what the register costs you.
Somebody to talk to
A conversation partner held to the words you already know. One unknown word per sentence turns practice into decoding, and the app knows your vocabulary exactly, which a human tutor has to guess at.
A story at your level
Not a level band. Two learners at B1 know different words, so a graded reader is always slightly wrong for everybody. This one is written from your own vocabulary.
The pattern in your mistakes
People rarely get things wrong at random. They get the same three things wrong, and nobody usually tells them which three, because nobody is watching that closely.
Mark what you wrote
The corrected sentence, the rule behind the change, and one more natural way to say the same thing. Writing is where a language stops being recognition, and it is what self-study normally cannot give you.
More practice on one rule
The curriculum gives every rule a few drills. Somebody who has failed all of them has run out of practice on the one thing they need practice on.
A better mnemonic
Three alternatives for a word that is not sticking, and a plain statement that your own hook beats all three even when it is worse, which is the one finding in this field nobody argues with.
Hooks for words that have none
The sound-alike engine does this with no key at all. A model does it better. Either way the result is labelled by where it came from and never counted as written.
Every one of these answers without a key.
Not a degraded error message: a real answer, built from the seven hundred written rules, the shipped stories, the word list and your own history. The offline answers are narrower, and the screen says which kind you are reading before you read it rather than after. What they never do is fail.
The rules we hold it to
What a model is not allowed to do here.
A language model will produce a confident sentence about any grammar point in any language, including the ones it is wrong about. That is a real problem for a teaching product and pretending otherwise would be the easy thing to do.
- Model output is labelled as model output everywhere a learner sees it
- Nothing the engine makes is ever counted with the hooks written for the library
- A generated drill whose answer is not among its own options is discarded rather than shown
- The tutor can be switched off entirely, per account, in settings
- With it on, it sends the word or question you asked about and the words you already know, never your name, your email or your review history
Start with ten words tonight
Thirty days free, no card. When the trial ends the account drops to the free limits and nothing is deleted.