How to practice Mandarin with an AI tutor (without treating it as the examiner)

It is good for forcing a full sentence when your deck only asks for a gloss. It is good for repeating one structure three times with one new constraint. It is good for hearing yourself say the fix out loud before the chat goes cold.

It is not a shortcut to fluency. It does not replace a live teacher. It is not Duolingo, Anki, Pleco, or HelloChinese. It is a text (or voice) partner with short memory, or, at best, with your level band.

This week, keep the job small. One structure: 因为…所以…, 想 / 要, or a counter dialogue. One measurable goal: three lines said aloud, not twenty chat screens.

If you already use ChatGPT voice, the mic is not the upgrade. The brief is. Without a level and without your error list, the model freestyles. Freestyle adds volume. It also adds false confidence.

A tutor that gets “I am at HSK 2” narrows the vocab. It still is not HSKK, the speaking exam separate from the written paper. Official papers and dates live on chinesetest.cn.

It also works as a warm-up before a real exchange. Five minutes on a closed prompt. Then you talk to a person. The tutor does not replace the human. It lowers panic on the first line.

It does not work for “review all of HSK 2” in one sitting. If the prompt opens too wide, the model changes topic every two turns. You get tired. You fix nothing.

The AI will miss tones. It invents a diacritic in pinyin (the official romanization for Mandarin) and writes it like a fact. You memorize it. A native ear does not.

It will miss measure words. It lets 两个苹果 through without 个, or it drops a book classifier on a fruit. If you never check, the error becomes habit.

It will miss polite register. It hands you 您 in a friend chat, or a blunt 喂 where courtesy belongs. A generic chat does not live your social context.

It will invent grammar. It glues particles because they “sound Chinese.” 了 appears or vanishes with no rule. You copy the pretty line. A graded text later contradicts you.

Hanzi (Chinese characters) can go wrong too: a rare variant, a near-homograph, or a gloss outside your band. The model is not a dictionary. It is a probability engine.

Tones deserve their own page: Mandarin tones for Spanish speakers. The rule here is enough: if the tutor flags a tone, ask; then verify outside the chat.

Another common miss: it “approves” a bad line because your brief was vague. You asked “correct me.” It gave you cheerleading. Without a criterion, the model is nice. Nice is not useful.

If you use voice, recognition fails too. It mishears your tone 3. It “fixes” a tone 2 you said right. Check a short native clip, not the chat’s confidence.

Ten minutes work if you close on time. Keep this shape.

Minute 0–1: one prompt. Level, structure, situation. Not “teach me Chinese.” Sample brief: “I am at HSK 2. I want one sentence with 因为 to say why I was late. Fix only grammar and tone. Do not invent HSK 4 vocab.”

Minute 1–3: you write or say your line. Do not polish it five times in silence. The visible error is the material.

Minute 3–5: one correction. One. If the tutor dumps a paragraph, ask: “Return only the fixed sentence and a one-line note.”

Minute 5–8: say the fix aloud twice. Record 20 seconds if you can. The chat does not replace your mouth.

Minute 8–10: one new constraint. Same structure, new subject, or no peeking at pinyin. Close. No “one more.”

That close is the product. Without it, the tutor becomes scroll. Scroll does not lock tone.

Prompt: 用「因为」说 why you arrived late. HSK 2. One sentence. Correct, then stop.

Weak learner line: 我迟到因为地铁.

Aspect is missing. The pause is missing. 因为 opens the reason; it should not stick to the verb with no clause. Push toward: 我迟到了，因为地铁太慢 (wǒ chídào le, yīnwèi dìtiě tài màn, I was late because the metro was too slow).

Why 了? It marks the relevant change: the lateness already happened. Without 了, 我迟到 can read more like a state or habit, depending on context. With 了, you report today’s fact.

Why does the reason follow 因为? Because 因为 opens the cause. In this short line, the cause is 地铁太慢. You can flip with 因为…所以…, but you do not need 所以 if the first clause already names the result (迟到了).

Next step, required: ask the tutor “Is the tone on 因为 right?” Then do not stop at its “yes.” Open a dictionary or native audio. 因 is first tone (yīn); 为 here is fourth (wèi) in the usual reading of the conjunction. If the model invents another pattern, the dictionary wins.

Say the good line aloud. Swap one piece: 因为下雨 (yīnwèi xià yǔ, because it was raining). Same frame. End the session.

That is output with a brake. It is not an oral exam. It is controlled repetition.

Next-day variant: same prompt, no peeking at the good line. If you produce 我迟到因为地铁 again, the work was not the chat. It was the mouth. Record. Compare with yesterday’s audio if you have it.

You can ask for a two-turn mini dialogue after the sentence. Only if you still have minutes. If not, close. A long dialogue is the enemy of a 10-minute session.

The numbers matter because the cap forces the short session.

Free: HSK 1–2, three exercises per feature per day, three tutor messages per day. Enough to test the method. Not enough for an afternoon of chat.

Standard: €5 a month (year and lifetime on /pricing). Twenty tutor messages a day. Still capped. The cap is a pedagogy feature wearing a cost limit.

Premium: €10 a month (year and lifetime again on pricing). No daily tutor cap. Memory of your level. It can generate extra readings at your band. There is a 7-day trial: cancel before it ends and you are not charged.

Why does Premium cost more? Model use has a real cost, and level memory saves you from repeating the brief every turn. That does not promise endless quality. The AI can still miss a tone. You pay for continuity and less friction, not for an examiner.

Public vocab inventory on the site: 12,432 words. Not the official paper. The paper and registration go to chinesetest.cn.

There are 204 graded readings and dialogues with audio on the site. Premium can generate more at your level. That is input. The tutor is output. Do not mix both in the same 10-minute window.

Standard is still a ceiling. Twenty messages cover several short sessions, not an endless chat the night before the exam. If you burn twenty in one thread, you lost the method.

Premium with no daily cap does not mean “keep going until it sounds perfect.” It means model cost no longer cuts you at three turns. You still cut at ten minutes.

Close the chat once you have said the fixed line aloud. The next click does not improve the tone. It only stretches the session.

Switch to SRS when tomorrow the right word will not come. The tutor gave you 迟到. If the gloss is gone tomorrow, that is what the card is for. The tutor is not your review queue.

Switch to listening when you can read the line and cannot catch it at normal speed. Chat lets you read. Audio lets you fail the rhythm. Failing the rhythm is the data.

Also close when you start bargaining with the model (“Are you sure it is fourth tone?”) more than once. That bargain is dictionary work, not tutor work. The tutor already did its job: it gave a hypothesis. You check outside.

If you collect three different fixes in one prompt, the session broke. Keep one. Park the rest for tomorrow. A brain will not lock a paragraph of nuance in one block.

When the official oral gets close, chat does not replace timed drills with human scoring or HSKK format. Use the AI to warm the mouth. Leave exam judgment to whoever scores for real, or to official material.

Also close if the tutor starts teaching lists. “Here are twenty synonyms for 因为.” That is a dictionary in disguise. You asked for one sentence. Pull it back: one line, one note, your voice.

If you end the session and cannot repeat the fixed line without looking at the screen, the session did not count. Say it once more. Then SRS or listening.

A typical week: three tutor sessions of 10 minutes on different days. The rest is queue or audio. The tutor is not the center of the plan. It is the output block.

If you want a tutor with level memory and no daily message cap, Premium’s trial lasts 7 days. Cancel before it ends and you are not charged. Live detail is on pricing.

Bring one prompt, one line of yours, one fix, and your voice. The examiner still lives on chinesetest.cn. The tutor only gets you talking today.

• One prompt, your sentence, the correction, say the correction aloud.
• Check a tone in a dictionary when the model sounds sure.
• Stop at 10 minutes even if messages remain.