Most people who try to practise a language with a chatbot do the same thing: type "let's chat in Spanish", get a polite stream of perfect sentences, and quit after three days. The model either corrects nothing, or it corrects everything and kills the conversation. The difference between a useful AI language learning conversation partner and a pleasant waste of time is almost entirely in the setup prompt.
This post gives you a setup prompt, three correction styles to pick from, a vocabulary loop that survives between sessions, and a plain list of what a chatbot can't do for you. All prompts here are untested templates unless I say otherwise; I did not run them on a live model for this post, and the example exchange is illustrative, written by me.
What should the setup prompt say?
A model left alone will pick a level, a topic and a correction policy for you. Decide those yourself. Here's a template:
You are my conversation partner for practising [LANGUAGE].
MY LEVEL: [e.g. A2, or "I know present tense and about 800 words"]
TOPIC TODAY: [e.g. ordering food, introducing my job, a phone call to book a room]
YOUR ROLE: [e.g. a waiter in a cafe in Madrid]
Rules:
1. Speak only in [LANGUAGE] unless I write "HELP" (then explain in English).
2. Keep each turn to 1-3 short sentences and end with a question, so I do most
of the talking.
3. Use vocabulary and grammar at my level; introduce at most 2 new words per
turn and put them in bold.
4. Correction style: [PICK ONE FROM BELOW].
5. If I write something you can't understand, say so as a real person would
("sorry, I didn't follow") instead of guessing what I meant.
Rule 5 is the one people forget. A model's default is to guess your meaning charitably, which means you never learn that your sentence was unintelligible. Making it ask for clarification is closer to real conversation.
If you want the same "stay in role, then give feedback at the end" structure for another skill, the mock interviewer prompts use it.
For your level, the CEFR labels (A1 to C2) are the common shorthand. I could not load the Council of Europe's own level descriptions on 2026-10-09 (the pages returned a 403 error), so I'm not quoting them here. Describing your level in plain terms, as in the template, works just as well in a prompt.
Which correction style should I choose?
Pick one per session. Mixing them is what makes the model erratic.
| Style | What it does | Good for | Cost |
|---|---|---|---|
| Recast | Replies with your sentence said correctly, without comment, and continues | Beginners who need flow | You may not notice the change |
| End-of-chat review | Says nothing until you type "DONE", then lists your errors with fixes | Building confidence, longer conversations | You repeat a mistake many times before seeing it |
| Immediate with a limit | Corrects only the single most important error per turn, in one line | Intermediate learners who want pressure | Slows the conversation |
Prompt wording for each, to paste as rule 4:
RECAST: When I make a mistake, reply naturally using the correct form of what I
meant. Do not point out the error or explain it.
END-OF-CHAT: Do not correct me during the conversation. When I type DONE, give
me a table: what I wrote | corrected version | one-line reason. Include only
errors that you are sure about, and mark any you are unsure about as "check".
ONE-ERROR: After each of my messages, add one line starting with "Fix:" that
corrects only the most important error (the one most likely to confuse a native
speaker). Ignore minor issues. If there's nothing important, write "Fix: none".
The "mark as check" and "only errors you are sure about" lines matter. They don't make the model reliable, but they make it surface its own doubt, which gives you something to verify.
What does a session look like?
This exchange is illustrative. I wrote it to show the shape of a session in Spanish at roughly A2, with the end-of-chat style. It isn't a captured transcript.
Partner: Buenas tardes. ¿Qué desea tomar?
You: Quiero un café con leche y un bocadillo de queso, por favor.
Partner: Muy bien. ¿Quiere el café caliente o frío?
You: Caliente. Yo soy muy cansado hoy.
Partner: Claro. Entonces un café con leche caliente. ¿Algo más?
You: DONE
| What I wrote | Corrected | Reason |
|----------------------------|----------------------------|-------------------------------------------------|
| Yo soy muy cansado hoy. | Estoy muy cansado hoy. | Tiredness is a temporary state: use estar. |
| Caliente. | (fine) | Short answers are normal in a cafe. |
The correction here (ser versus estar for a temporary state) is the sort of explanation you can check in any grammar reference, and it's a standard point. The next section is about the corrections you should check.
How do I verify corrections and avoid learning wrong things?
A chatbot can be wrong in a way that sounds exactly like a teacher. Build in checks:
- Ask for the rule plus two examples, then test them. If the two examples don't obey the rule it just gave, the rule is suspect. Cheap and effective.
- Look up anything you plan to memorise. A bilingual dictionary entry or a learner's grammar reference settles most disputes in under a minute. The same habit applies to any AI answer; the fact-checking checklist covers the general version, and the hallucinations lesson explains why confident wrong answers happen.
- Ask it to disagree. Paste a corrected sentence into a fresh chat and ask, "Is this natural, or would a native speaker say it differently? Give 2 alternatives." Different phrasings can reveal that the first correction was grammatical but stiff.
- Keep a "check with a human" list. Once a week, take the five explanations you were least sure of to a tutor, a language exchange partner or a forum for that language.
- Beware of the model agreeing with you. If you push back on a correct correction, a model may cave (I did not measure how often). Test it: ask a question where you know the answer, make a wrong objection, and see whether it holds its position. It's a quick measure of how much to trust a given session.
What can't it do?
State the limits plainly, because they shape how you use it.
- Accent and pronunciation. A text chatbot hears nothing. Voice modes transcribe your speech first, and transcription can quietly fix your mistakes or invent errors you didn't make. A transcript is not a pronunciation score. For accent work, record yourself and compare with a native recording, or work with a person.
- Cultural nuance and register. Whether a phrase is rude, old-fashioned, too formal or regionally wrong depends on country, age and context. Models tend to produce a textbook-neutral register. Ask explicitly ("which country's usage is this?"), and don't treat the answer as final.
- Regional and low-resource languages. Quality drops for languages with less written data, and dialect forms are often flattened to a standard. If you study one of those, see the notes in translate with AI: regional languages quality checklist and prompting in Hindi and Indian languages. Verify with a speaker.
- Real conversation pressure. The model waits for you, speaks clearly and never interrupts. Real people don't. Use the chatbot to build volume and confidence, then talk to humans.
- Memory between sessions. A new chat starts blank unless you've enabled a memory feature, and even then you can't rely on it to track your progress. Keep your own log (next section).
How do I build a vocabulary loop that sticks?
New words in a chat vanish when the chat does. A loop you control looks like this:
- At the end of each session ask: "List the new words and phrases I met today as a table: word | meaning | one example sentence from our chat. Use only words that appeared." Copy it into your own notes with the date.
- Next session, paste the last ten entries and add rule 6 to the setup prompt: "Try to use these words naturally in your questions, and note when I use them correctly."
- Once a week, turn the list into flashcards. The Anki flashcard workflow post has a script that converts a notes list into an import file. Put the sentence from the chat on the card, not just the word.
- Spot-check the translations. Single-word meanings from a model are often right, but words with several senses need a dictionary look.
A recap prompt for the end of a session:
Summarise today's session: (1) 5 new words with meanings, (2) my 3 most
frequent mistakes with the correct form, (3) one grammar point to study next.
Say "unsure" next to anything you are not confident about.
What to do about privacy and copyright
Don't paste personal details, passwords or other people's messages into a practice chat. If you use learning materials such as a textbook's dialogues, use short excerpts you're entitled to use, or write your own scenarios. A scenario you invent works as well as a copied one.
A one-week starter plan
- Day 1-2: Typed role-play with the recast style, 10 minutes. Topic: a task from your real life (ordering, introductions).
- Day 3: Same topic, end-of-chat style. Compare the error table with what you remembered.
- Day 4: Pick one repeated error, ask for the rule and two examples, then verify it in a reference.
- Day 5: One-error style, a harder scenario (a problem to resolve, such as a complaint or a change of booking).
- Day 6: Speak it aloud if you can, record yourself, and compare with a native recording.
- Day 7: Take your five doubtful corrections to a human, and turn your word list into cards.
The honest summary: a chatbot gives you cheap, patient repetition, and you supply the judgement about what's correct and natural. If you keep the correction style fixed, keep a log and check what you plan to memorise, it's a good use of ten minutes a day.
