A tutor that has read your mistakes
A tutor that knows your level, your vocabulary and the mistakes you made this week.
It is not a general chatbot with a Chinese prompt bolted on. Every answer is grounded in the same reviewed records the rest of the platform is built from, it is told your level and your recent mistakes, and it says what it used.
10 tutor messages a day on the free plan, in explain and correct mode.
What is the difference between 了 and 过?
了 marks that an action is completed — usually a specific event: 我吃了饭 (I ate / I have eaten). 过 marks that you have had the experience at some point: 我吃过北京烤鸭 (I have tried Peking duck before). The test is whether you care about *finishing* something, or about *having tried* it. With 过 the experience is often in the past and no longer relevant now; with 了 the completion often still matters.
Modes
Seven modes, because “ask me anything” is not a teaching method
The mode you pick changes what the tutor retrieves, how much it says, whether it answers in Chinese, and whether it corrects you immediately or waits until you finish.
Explain
Grammar and vocabulary explained at your level
Ask why a sentence works, or what separates two words you keep mixing up. The answer is written with the vocabulary you have already met, in your interface language, and ends with the smallest possible next step rather than a wall of theory.
Grounded in: Grammar points, vocabulary senses, example sentences, confusable pairs.
Practice
Targeted questions built from your weak items
The tutor builds questions from the items you actually failed this week and the modes you failed them in — a word you can read but cannot hear becomes a listening question, not another reading card.
Grounded in: Your review history, the item bank, grammar exercises.
Role-play
A scenario played out in Chinese, with corrections
A scenario is played out in Chinese at your level: registering at a university office, seeing a doctor, negotiating a deposit. Corrections come after each turn so the conversation is not shredded mid-sentence.
Grounded in: Survival Chinese scenarios, dialogues, level vocabulary lists.
Correct my sentence
Paste Chinese you wrote and get it corrected
Paste Chinese you wrote. You get the corrected sentence, a diff of what changed, the rule you broke with a link to it, and a more natural alternative when the grammar was fine but the phrasing was not.
Grounded in: Grammar points, common-mistake records, collocations.
Exam coach
Strategy and practice for your target examination
Strategy and timing for the paper you are sitting, built from your own mock results. It coaches technique — it does not state the rules of the examination from memory, and any structural claim is traced to a source record with a verification date.
Grounded in: Your mock attempts, gap analyses, verified examination source records.
Survival Chinese
Phrases for a real situation, right now
For the situation you are in right now. Short, usable phrases with pinyin and audio, plus the cultural context that makes them land, and an offline phrase card you can keep on your phone.
Grounded in: Survival Chinese scenarios and their cultural notes.
Daily conversation
Everyday conversation at your level
An unhurried conversation about ordinary things, held at the edge of what you can handle. New words are introduced deliberately, glossed once, and offered to your review queue at the end.
Grounded in: Level vocabulary, graded stories, your mastered-word list.
Retrieval
It looks the answer up before it writes it
Every question runs through the same pipeline. Nothing is answered from the model’s own recollection of Chinese, and nothing that has not been reviewed by a human editor can be retrieved.
Your question is read together with your profile
Your curriculum track, current level, target level, examination date, the words you have mastered and the mistakes you made this week are part of the request. This is what lets the tutor answer in vocabulary you can actually read.
Reviewed records are retrieved
The question is matched against published vocabulary, characters, radicals, grammar points, example sentences, lessons, stories, scenarios and articles. Draft records, records still in review and AI-drafted records that no editor has approved are excluded from the index entirely.
The answer is written against those records
The retrieved passages are given to the model as the material it must answer from, with an instruction to say it does not know rather than to fill a gap. Examination structure, scoring and dates may only come from a source record with a verification date.
The sources are shown to you
Each answer lists the records it used, as links you can open. If an answer cites nothing, that is visible — and it is your signal to treat it as a suggestion rather than a fact.
The exchange becomes study material
A correction can be turned into a micro-lesson, a quiz, or new cards in your review queue in one action, so the conversation feeds the scheduler instead of evaporating.
What it can retrieve, right now
2882 published records in this environment. Only these can be cited.
- Vocabulary
- 752
- Characters
- 140
- Radicals
- 214
- Grammar
- 44
- Sentences
- 1,638
- Lessons
- 32
- Stories
- 22
- Scenarios
- 20
- Articles
- 20
Counts are of records with the status “published”. Drafts and unreviewed AI drafts are not included and cannot be retrieved.
Which model answers you
DEV MOCKThis environment is running the built-in local mock model. Its replies are deterministic scaffolding for testing the interface — they are not teaching material, and every surface that shows them carries a DEV MOCK badge.
Limits
What the tutor will not do
These are product rules, not aspirations. Each one is enforced in the prompt, in what the retriever is allowed to return, and in what the interface renders.
It will not state official examination rules from memory
The tutor does not state examination rules from memory. Anything about examination structure or scoring comes from a reviewed source record, with its verification date.
Ask it how many words a level requires and it will point you at the examination authority and at the source record behind any figure the platform shows — including the date that figure was last checked.
It will not pretend to be a human teacher
It identifies itself as software in the first message of every conversation and whenever asked. It does not claim credentials, a name, a life, or feelings about your progress, and it will not role-play as a real, identifiable person.
In role-play mode it plays a role — a registration clerk, a pharmacist — and the interface still labels the conversation as a practice scenario with a machine.
It will not present generated text as verified
Anything a model produced is labelled as such. Generated text never enters the dictionary, the grammar library, the lesson bank or a mock examination, and it is never shown with the styling used for reviewed content.
A model-drafted record can only reach the published library after a human editor has reviewed and approved it — the platform cannot publish an AI draft automatically.
And three smaller ones
- It will not grade an official examination or predict a score you will get. It gives a readiness estimate built from your own results, and labels it an estimate.
- It will not answer a question that has nothing to do with learning Chinese by pretending to be a general assistant. It says what it is for and stops.
- It will not silently rewrite your Chinese. A correction always shows what you wrote, what to change, and which rule it broke.
Privacy
What happens to what you type
A tutor that reads your mistakes needs your mistakes. Here is the whole arrangement, in the order it matters.
What is sent
Your message, the retrieved records, and a profile summary: level, curriculum track, target, examination date if you set one, recent error types and a sample of weak items. Your name, email address and payment details are never part of a model request.
What is stored
Conversations are stored on your account so you can reopen them, and so a correction can be turned into review cards later. You can delete a single conversation or all of them from the tutor history, and deletion is immediate rather than a flag.
Training
Your conversations are not used to train models. When a hosted provider is configured, requests are sent under an agreement that excludes training on submitted data; the provider in use is named on this page and in the privacy policy.
Voice
Speaking practice follows the same rule as the pronunciation lab: by default the recording is analysed and discarded, and only the derived scores are kept. Retention is opt-in, per account, and every retained recording can be played back and deleted individually.
Children
Where a school enrols learners under 18, tutor conversations belong to the learner. A teacher sees completion and weakness reports, not the transcript.
The full position is in the Privacy policy
Frequently asked questions
Is the tutor a replacement for a teacher?
No, and it does not claim to be. It is available at midnight, it has read every mistake you made this week, and it never gets bored of the same question — but it cannot watch your mouth while you speak, it cannot tell that you are discouraged, and it has no judgement about your life. Teachers and schools use it as the drilling layer between lessons, not instead of them.
What happens if it does not know?
It says so. The retrieval step either returns reviewed records or it does not, and when it does not the tutor is instructed to say that the platform has nothing published on the point and to suggest where to look, rather than to produce a plausible answer.
Can it tell me the rules of the HSK examination?
It will not state them from memory. Where the platform holds a source record for a structural claim, the tutor quotes it and shows the publisher and the date it was last verified. Where it does not, it directs you to the examination authority or your registered test centre. HSK Guide is independent of the examination authorities and does not speak for them.
Which language does it answer in?
Explanations come in your interface language — English, French, Arabic or Chinese. Chinese examples, corrections and role-play turns are in Chinese, at or just above your current level, with pinyin available and audio on every line.
Does it work on the free plan?
Yes, with a daily message allowance and two of the seven modes: explain and correct my sentence. Pro removes the allowance and opens the remaining five modes.