How to use an AI tutor for Chinese without it teaching you the wrong words
Ask any chat model to practice Chinese with you and it will happily produce sentences full of words you have never met. It has no idea what you know, so it guesses, and it guesses high. The fix is small and it changes the tool completely: hand the model an explicit list of your vocabulary and grammar, and tell it that everything outside the list is unknown to you.
The problem, precisely
A learner told us in August 2026: "I've tried actually using voice LLMs — but the LLM doesn't know what my current vocab is, or what grammatical concepts I've learnt." That is the whole failure. A tutor pitched above your level produces output you cannot use, and worse, you cannot tell which parts were supposed to be new. Level labels do not fix it, because "I'm HSK 3" is an abstraction the model has to interpret, and its interpretation of HSK 3 is not your HSK 3.
What to give it instead
A list. Four parts, in this order:
- An instruction block. Build every Chinese sentence from the vocabulary below; if you need a word that is not on the list, say so in English rather than slipping it in.
- The sentence patterns you have been taught, each with one worked example. This is the grammar ceiling, and it is what stops the model from answering a beginner in 把 sentences.
- The vocabulary you have met, with pinyin and meaning. Weakest first, so a model that reads only the top of a long paste still reads the part that matters.
- What is shaky. Items you get wrong repeatedly are the ones worth drilling; without this the tutor spreads its attention evenly, which is the one distribution you do not want.
Then add the rule that makes the whole thing work: everything not on these lists is unknown to me — treat the lists as the ceiling.
The one-click version
GraphChinese exports exactly that file. On the 说 Speak tab there is a study snapshot: copy it to the clipboard, download it as markdown, or open it in a tab. It contains your working level, every topic you have met with a strength estimate that decays between reviews, every sentence pattern with its example sentences, every distinct word with pinyin, meaning, level and strength, and the tutor instruction block at the top. Weakest items come first throughout.
Why strength rather than a checklist: a word you met in March and have not reviewed since is not the same as a word you drilled yesterday, and a tutor that treats them alike wastes your session. The snapshot carries a 0–100 estimate of how well each item is held right now, so "drill the low numbers" is a one-line instruction.
The snapshot is built from your own study record, so it is available to members. If you are not one, the next section builds a smaller version by hand.
Rolling your own in ten minutes
- Take the vocabulary list for your level. The complete HSK lists are free here with pinyin and English: HSK 1, HSK 2, HSK 3, HSK 4, HSK 5, HSK 6.
- Cut what you do not actually know. This is the step that matters and the step people skip. A list you half-know produces a tutor pitched at the half you do not.
- Add ten patterns you can use, with one example each. Ten real ones beat forty aspirational ones.
- Paste it under an instruction block and keep it in a note. Update it monthly; it is the same file every session.
What to use it for
- Production drills. English prompt, you answer in Chinese, it corrects. This is the direction that is hard to practice alone and the direction every exam and conversation asks for.
- Sentence mining at your level. Ask for ten sentences using three specific weak words, then read them aloud.
- Explaining your own errors. Paste a sentence you wrote and ask what a native speaker would say instead, and why.
- Voice practice. A voice assistant with the snapshot loaded is a patient partner at your exact level, which is a thing that does not otherwise exist at 6am.
Ready-made prompts for each of these are on the Chinese tutor prompts page.
What it cannot do
It is not a conversation. A model will not misunderstand you, will not get bored, will not change the subject to its grandchildren, and will not make the social stakes real — and those four things are what a beginner needs from a person. Use the model for repetition, and use people for conversation. How to run the human half.
It is also not a curriculum. A model will drill whatever you paste, in whatever order you ask, which means it can only be as well sequenced as the list you gave it. Deciding what to learn next is the part that has to come from somewhere else.
Frequently asked questions
Why is an AI tutor bad at teaching Chinese by default?
Because it does not know what you know. Asked to practice Chinese, a model produces sentences at the level it guesses you are, which is usually far above your actual vocabulary. A tutor that invents HSK 5 words at an HSK 2 learner is worse than no tutor, because you cannot tell which parts you were supposed to understand.
How do you tell ChatGPT what Chinese you already know?
Paste a list. The model follows an explicit vocabulary and grammar list far more reliably than a level label, because "HSK 3" is an abstraction it has to guess at while a list of 600 words is a constraint it can check against. Tell it plainly that anything not on the list is unknown to you.
What should be in the file you give an AI tutor?
The words you have met with their pinyin and meaning, the sentence patterns you have been taught with one example each, which items are shaky, and an instruction to build every sentence from that list only. Adding how well you hold each item is what lets the tutor drill the weak ones instead of the whole list.
Can an AI tutor replace a human conversation partner?
No, and it is very good at the part a human is expensive for: unlimited patient repetition of material you choose, at any hour. Use it for drilling production and use a person for real conversation, where the unpredictability and the social pressure are the point.
Does GraphChinese export your data for this?
Yes. The study snapshot is a plain markdown file listing every topic and word you have met, weakest first, with instructions for a tutor at the top. You copy it or download it from the Speak tab and paste it into any chat or voice assistant. It is available to members, because it is built from your own study record.
Your study record, as a file a model can read. GraphChinese exports every topic and word you have met — weakest first, with strengths and tutor instructions — from the Speak tab, ready to paste into any assistant. The whole course, HSK 0 through HSK 6, is $49 once for lifetime access.
Start at your true level