Foreign Languages in High School: Conversation with AI Instead of Expensive Tutoring

You don't learn a language by reading about it. You learn it by using it — speaking, messing up, getting corrected, and trying again. That's exactly what's in short supply at school: in a class of thirty, each student gets a couple of minutes of speaking time per lesson, and not everyone can afford a private tutor for an hour of speaking practice.

AI fills that gap better than you'd expect: it's a partner you can talk to anytime, about anything, for as long as you want, and it never winces at a mistake. It won't replace a teacher or time spent living abroad — and it does a few things badly, which this guide covers openly. But for what most students are missing the most — sheer speaking mileage and instant feedback on their own writing — it's a tool that changes outcomes.

This guide walks from setup to daily routine. You don't have to go through it all at once: feel free to start at Phase 2 with one conversation tonight. Add the rest as you feel the need for it.

A typical scenario

Kristýna is a sophomore, and her English is “just barely a C.” She understands most of what she hears, handles the grammar fine in her workbook, but when she has to say something in class, five words come out and then silence. She tried tutoring for two months, an hour a week — and nothing happened between sessions.

She changed one thing: every day for fifteen minutes, she talks to AI in voice mode, usually on the walk home from school. She set it up once — level B1, correct only after the conversation, not during it, speak more slowly, don't switch to her native language. She pulls topics from her final-exam speaking topic list, so she's simultaneously prepping for her oral exam two years ahead of time.

After every conversation, she has it list the mistakes she made and three phrases she was missing. She adds those phrases to flashcards. She writes her essays herself, but has them checked — under a rule that they can't be rewritten, only explained: what's wrong and why. After four months, the difference is audible: not in grammar, which was always decent, but in the fact that sentences actually come out of her mouth at all. The rest of this guide is the process that gets you there.

Phase 1: setting up your partner

Most people try speaking to AI in a foreign language, get an answer full of long, complicated sentences, understand half of it, and conclude it doesn't work. The problem wasn't the tool — it's that nobody told it who it was talking to.

Figure out your level, so you know where to start

The A2, B1, and B2 levels aren't bureaucracy — they're practical information: they determine how complex the sentences your partner should be offering you are. Roughly speaking, A2 means getting by in everyday situations with short sentences, B1 means holding a conversation on familiar topics and describing an experience or opinion, B2 means speaking fluently even about more abstract things and understanding longer continuous speech. For a foreign-language exam, most students aim for around a B1, or a B2 for a higher-level exam — check the specific requirements against your official exam syllabus, not a chat.

I want to estimate my level in [English] on the A2/B1/B2 scale.

Do it this way:
1. Ask me 8 questions in [English], from simple to more complex —
   ones that reveal my vocabulary range and confidence with tenses.
   Ask one at a time and wait for my answer.
2. Don't correct or comment on anything along the way, just keep
   asking.
3. At the end, tell me in [my native language]: an estimated level,
   three specific things I do well, and three things holding me
   back (with an example from my answers for each)
4. Recommend what level I should hold conversations at to keep
   improving — not so it's comfortable, but not so I can't
   understand either

Treat this as an estimate, not a certificate. Start with the first
question.

This returns a rough placement and, more importantly, a list of specific weaknesses that becomes your assignment for the next month. No pretense of certification here — it's an estimate from eight sentences, not an exam.

Instructions you set once

The best investment in this whole guide is one well-written instruction you won't have to repeat in every conversation. Save it in a Project (persistent context in claude.ai), in your custom instructions, or at least keep it in your notes so you can paste it in.

From now on, you're my conversation partner for [English]. Stick to
these rules in every conversation until I say otherwise:

MY LEVEL: [B1]. Use vocabulary and grammar at this level, and
occasionally throw in one harder word, then immediately reuse it
in another sentence so I can figure it out from context.

HOW TO TALK: short sentences, natural pace, no long complex
sentences. Speak in [English], and only switch to [my native
language] if I explicitly ask you to, or if I fail to understand
three times in a row.

CORRECTIONS: do NOT correct or interrupt me during the
conversation, even for major mistakes. The conversation needs to
keep flowing. If you don't understand me, rephrase instead of
explaining.

ASK QUESTIONS: always end your response with a question, so I
have something to respond to. If I answer in one sentence, follow
up for detail.

AFTER THE CONVERSATION: only once I say "done," give me the
analysis — in [my native language], following the rules I'll
send you.

Understood? Reply in one sentence and start with the topic
[topic].

This returns a partner you can actually talk to. The paragraph about corrections is the most important one — the model has a strong tendency to jump in and correct you, which is exactly what kills fluency. If it forgets, just say “don't interrupt me, save corrections for the end.”

Why correct only after the conversation

It sounds like a minor detail, but it's a fundamental difference. Being corrected mid-sentence makes you think about grammar instead of content, slows your speech down, and eventually you'd rather stay quiet. The goal of conversation is fluency — the ability to produce a sentence without a long pause. Accuracy gets handled afterward, calmly, when you're looking at a transcript of mistakes.

Split it into two separate activities in your head: speaking (must never be interrupted) and review (happens afterward). Mix the two, and you do both badly.

Phase 2: conversation that actually gets you somewhere

What fifteen minutes looks like

A short conversation with a clear task beats half an hour of free chatting. A proven structure: one minute to warm up (what happened today), ten minutes on a topic or scenario, three minutes to summarize, where you recap what you talked about in your own words. That last part is the most valuable and the most often skipped — summarizing is exactly when you get pushed to say something you can't quite manage yet, and that's where the learning happens.

You can do this by voice or by typing. Voice mode is better for languages, because it also trains your mouth to move faster than your head; more on it in AI voice mode. Typing works well when you're on a crowded bus, or when you want a transcript right away.

Scenarios instead of free chat

“Let's talk about travel” is a weaker prompt than “you're an airport employee, I missed my flight, and I need to sort it out.” A scenario forces you to use specific language, gives the conversation direction, and you can't escape into sentences you already have memorized.

Let's do a roleplay in [English]. My level is [B1].

SITUATION: [e.g. I'm a hotel guest, there's no hot water in my
room, and I want it fixed; you're the receptionist, who first
offers a solution that doesn't work for me]

RULES:
- stay in character, don't correct or explain anything along
  the way
- don't make it easy: don't agree with everything right away,
  make me argue my case
- talk naturally, including common phrases a native speaker
  would actually use
- if I get stuck and type "help," give me one word of a hint
  and keep going in character

When I say "done," drop the roleplay and give me the analysis.
You start, with the first line.

This returns a dialogue that goes somewhere you don't expect, which is the point. If the scenario is too easy, add a complication (“the receptionist insists it's not their problem”). A stock of reusable scenarios with variations is worth more than a list of topics.

Topics from the exam speaking list

Once you're already talking, talk about what you'll actually need for the exam. Foreign-language exam speaking topics tend to be very general (family, housing, travel, education, health, culture, life in countries where the language is spoken), so you can cover them starting in your first year, and hit ordinary conversation practice at the same time.

Prepare a set of conversation prompts for the exam topic [topic],
at level [B1].

I want:
- 5 questions I might expect on an oral exam, ranging from
  descriptive to opinion-based
- for each, 6 useful phrases or expressions that fit it (not
  isolated words, whole phrases)
- 3 roleplay scenarios on this topic where I have to react,
  not just narrate
- 5 questions I'm afraid of, because I wouldn't have much to say
  even in my own language — with a hint for each on how to talk
  around it politely and fluently

Write the prompts in [English], phrases with a [native-language]
translation, notes in [my native language].

This returns material for several weeks of conversations. The last point is a practical trick for the oral exam: learning a handful of phrases for buying time or redirecting a question helps more than another twenty vocabulary words. This connects to the overall exam prep guide, the year-long plan for seniors.

Analysis, only after the conversation

This is where conversation turns into learning. If you type, you already have a transcript; for voice, ask for a summary of the conversation, or go through the transcript the app shows you.

End of conversation. Now give me the analysis, in [my native
language], in this order:

1. MISTAKES I made — only the ones that recurred or interfered
   with understanding. No more than 6. For each: what I said, how
   it should be, and ONE sentence on why (what rule is behind it).
2. NATIVE-LANGUAGE PATTERNS IN [ENGLISH]: sentences I clearly
   translated directly from [my native language] that a native
   speaker wouldn't phrase that way — and how it's normally said.
3. WHAT I WAS MISSING: 5 words or phrases I needed during the
   conversation and didn't have — each in the full sentence from
   our conversation.
4. WHAT WENT WELL: two things I said well.
5. One topic for next time that will push me to use what I just
   learned.

Don't correct minor things and don't list everything — I want
six mistakes that matter, not forty.

This returns an overview you can process in five minutes. Limiting the mistake count is deliberate: nobody remembers a list of forty corrections, and it's just demoralizing. Save point 3 — it's the best vocabulary source you have, because these are words you needed yourself.

Phase 3: vocabulary in context, not lists

Why a list of thirty words doesn't work

A “word — translation” column gets learned for Friday's quiz and is gone by Sunday. The reason is simple: a word with no context has nothing to attach to in your memory, and worse, you're training the wrong skill — recognizing a word when you see it, instead of recalling it when you need to say it.

The opposite approach works: a word in a full sentence, ideally one that's actually about you. Language gets stored in phrases, not individual words.

Here's a list of words I need to learn from [textbook, lesson X]:

[paste the list]

Turn them into material I can actually study from:
- for each word, write 2 sentences in [English] at level [B1] that
  use it naturally (not artificial textbook sentences)
- at least one of those sentences should be about ordinary
  high-school life
- for words with multiple meanings, show the difference with
  a pair of sentences
- for each word, note what it typically pairs with (common
  collocations, prepositions, verbs)
- flag words that get confused with a similar-looking word in
  [my native language], or that mean something different than
  they look like they do

Finally, write a short story (max 150 words) that uses as many
of these words together as possible.

This returns a set you can genuinely study from. The most valuable part tends to be the typical collocations — those are exactly what separates “technically correct” from “no native speaker would say it that way.” Read the closing story out loud — it's painless repetition.

Turn the material into flashcards

Words you simply need to know cold belong in spaced-repetition flashcards. The full process is covered in flashcards with AI; for languages, two things are specific: a card should contain a full sentence with the word blanked out (not an isolated word), and the learning direction should mainly go from your native language into the foreign one, since that trains recall, not recognition.

Turn these sentences and words into Anki flashcards:

[paste sentences and words, straight from the conversation
analysis if you like]

Format: two columns separated by a semicolon, one card per line,
no header, no numbering.

Rules:
- on the front, put the native-language sentence, or the sentence
  with a blank in the form "I ___ to the doctor yesterday. (go)"
- on the back, the correct full sentence, not just the missing word
- one card = one thing to remember, never a list
- for irregular verbs, make a separate card for the forms
- skip words that are identical in [my native language] and the
  foreign language

Return only those lines, nothing else.

This returns a file you can import directly. Look it over before importing — with fill-in-the-blank sentences, more than one answer sometimes fits, and it's better to rephrase a card like that than argue with it every morning.

A gaps log

The best vocabulary list isn't from a textbook — it's from your own gaps. Whenever you're in a conversation or writing and hit something you wanted to say but couldn't, jot it down in your native language in one file. Once a week, look up how to say those things and turn them into flashcards. In a month you'll have a vocabulary set built exactly around what you actually talk about.

Phase 4: grammar explained in your native language

Grammar is worth having explained in your native language — not because you're not up to the foreign language, but because the explanation needs to be fast and unambiguous. When you're still not clicking with the present perfect for the third time in your workbook, three minutes of explanation in your native language with examples solves more than another exercise.

I don't get [a grammar point, e.g. the difference between the
present perfect and simple past] in [English]. I'm [a high school
student], level [B1].

Explain it to me in [my native language] like this:
1. What this feature is actually for — one sentence, no jargon
2. The basic rule for when to use it and when not to
3. Five pairs of sentences that differ only in this feature, with
   a [native-language] translation of both — I want to see how
   the meaning changes
4. What specifically trips up speakers of [my native language]:
   why we struggle with it and which construction in our own
   language leads us to the wrong form
5. A simple mental test I can run through when I'm not sure which
   option to use

Don't give me a table of all the forms, I have plenty of those in
my textbook. I want to understand when to use what.

This returns an explanation focused on usage rather than forms. Point 4 is usually the one that makes it click — a lot of mistakes come from translating a native-language construction word for word. Then verify the rule against your textbook: on edge cases, models sometimes oversimplify or invent a rule that sounds logical and isn't real.

An exercise on one specific feature

Prepare an exercise on [grammar point] for level [B1].

- 15 fill-in-the-blank sentences where the feature genuinely
  needs to be distinguished (no sentences where either option
  would work)
- make the sentences about ordinary situations, not abstract
  examples
- increase the difficulty gradually, make the last three tricky
- separately, give 5 sentences in [my native language] to
  translate, so I have to produce it myself instead of just
  picking
- put the answers ALL THE WAY AT THE END, separated out, not
  next to individual sentences
- for each answer, add one sentence on why, especially for
  the tricky ones

After I check my answers, tell me whether I should keep
practicing this feature or move on to the next one.

This returns an exercise you can print out or fill in on paper. The condition “no sentences where either option would work” matters — without it you get questions you end up arguing with the answer key over.

Phase 5: writing — correct and explain, don't rewrite

Writing is where the temptation is greatest, and so is the cost. Having AI write your essay for you is the fastest way to hand in a piece of writing on your exam or a test that's two levels worse than what you've actually been “writing” so far. And most importantly: a corrected sentence you wrote yourself teaches you something; a finished text someone else wrote teaches you nothing.

The rule for this whole phase: you write the text, AI corrects and explains it, but never rewrites it. The difference comes down to what you ask for — and it has to be spelled out explicitly in the prompt, because a model's default behavior is “here's an improved version.”

I wrote an essay in [English]. The prompt was: [prompt, including
the required text type and length]. My level is [B1].

MY TEXT:
[paste the text exactly as you wrote it, mistakes included]

DON'T REWRITE IT FOR ME. I want feedback, not a finished text.
Do this:
1. List mistakes in a table: what I wrote | correct version |
   type of mistake (grammar / vocabulary / spelling / translated
   directly from my native language) | one-sentence explanation
   in [my native language]
2. Flag the 3 most serious mistakes — the ones that would hurt
   me most on an exam — and say why
3. Assess content and structure: did I answer the prompt, do
   I have an intro and conclusion, does the text stay on topic,
   are the length and text type correct
4. Write 5 phrases that would elevate the text — but just list
   them, don't insert them into the text
5. Tell me what to fix myself, and I'll send it back to you

Don't rewrite anything for me, not even one sentence.

This returns a list of mistakes with explanations you can actually work with. If the model still tacks on a “corrected version” at the end anyway, ignore it and tell it not to do that next time. The mistake type “translated directly from my native language” is the most common one for non-native students and the least visible — in a teacher's grading it shows up as “unnatural phrasing.”

Round two: comparing versions

Here's my second version of the text, after corrections:

[paste the corrected text]

Compare it to the previous version and tell me:
1. Which mistakes I fixed correctly
2. Which ones I fixed wrong or only halfway — and what's still
   off about them
3. Which ones I missed entirely
4. Whether a fix broke something (a sentence that was better
   before)
5. One thing to focus on for my next text

Again: don't rewrite the text. And don't hold back, I want to
know.

This returns a check that's more valuable than the first round of feedback — it reveals which explanations you actually understood. Point 4 happens more often than you'd expect: it's easy to break a sentence that was already fine while “fixing” it.

Formal emails and letters

Exam writing tasks and regular tests often ask for a letter, email, or complaint, where register — how formal the text is — gets graded too. This is something you can drill quickly, since it rests on a handful of set phrases.

I'm writing a [formal email / complaint / request] in [English]
about [topic], level [B1], length [120 words].

First give me material to work from, then I'll write it myself:
1. An outline: what goes in which paragraph
2. Set phrases for each part (greeting, reason for writing, main
   point, request, closing) — for each, note whether it's formal,
   neutral, or too informal
3. Five expressions that don't belong in a formal text, even
   though we use them casually all the time, and what to replace
   them with
4. What to watch out for specifically in [English] (punctuation
   after the greeting, abbreviations, introductory phrases)

Don't write the text for me, I'll write it myself and send it
to you for review.

This returns a toolkit you assemble the text from yourself. Save the phrases — formal texts recycle them endlessly, and it's one of the few places where memorizing genuinely pays off.

Phase 6: listening and pronunciation — and where AI hits its limits

Time to be honest here: there are two parts of language learning AI only handles partially, and it's better to know that up front than to rely on something that doesn't actually work.

Pronunciation: what AI can and can't tell

Voice mode first converts your speech to text, and only then works with that text. That has two consequences. First, pronunciation feedback is unreliable — when the model tells you “your pronunciation was good,” that doesn't carry much weight, because it's judging a transcript, not a sound. Second, the reverse signal is actually useful: if the transcript repeatedly shows a different word than what you said, that's a real clue that you're pronouncing that word in a way even a machine can't understand.

What to do about it in practice: for pronunciation, use a dictionary's phonetic transcription and native-speaker recordings, not a model's judgment. And practice shadowing — play a short recording and speak along with it, trying to copy the rhythm and intonation. Two minutes a day, and it's the most effective pronunciation exercise you can do on your own.

Listening: get your material from real people

AI's voice is synthetic and speaks more slowly and cleanly than real people do. Anyone who trains their listening only on it is in for an unpleasant surprise on an exam or abroad. Train listening on real material: podcasts, shows with subtitles in the target language, news for language learners, videos. AI is useful as prep and review around that material.

I'm going to listen to a [video / podcast] about [topic] in
[English], [8] minutes long, my level is [B1].

BEFORE LISTENING, prepare:
- 10 words and phrases that are almost certain to come up in
  material like this, with a translation and pronunciation
- 5 questions to listen for the answer to
- a heads-up on what's likely to be hardest (speed, accent,
  specialized vocabulary)

AFTER LISTENING, ask me those 5 questions and let me answer in
my own words. Then assess what I didn't understand, and whether
it was mostly vocabulary or the speed.

Don't tell me what the material is about — I want to hear that
for myself.

This returns prep that turns passive background listening into an exercise. The last line matters: the model can't actually play a specific video, and would happily describe it based on a guess from the title, which is worthless material.

What AI won't replace

A live partner who looks you in the eye and reacts to how you come across. Time spent among people who actually speak the language. A teacher who's known you for two years and knows exactly where you get stuck. Think of AI as a flight simulator between lessons — one you can log a lot of hours on, but it doesn't replace the actual flight.

Phase 7: a fifteen-minute daily routine

Language is a skill where one hour a week is worse than fifteen minutes every day. Consistency is what matters — which is exactly why fifteen minutes is deliberately small: it has to fit even on a day when nothing else does.

A proven weekly split: three days of conversation (fifteen minutes, a scenario or an exam topic, followed by a short review), two days of writing or grammar (one text or one exercise on one feature), one day of listening (one video with prep and review), one day off. Flashcards every day, but count those separately — it's five minutes on the bus in the morning.

Build me a fifteen-minute daily routine for [English], level [B1],
goal [improve speaking and prep for my final exam in two years].

My day looks like this: [when I have time, when I'm commuting,
when I have clubs/activities].

Break the week down by day:
- what specifically I do that day (an activity, not "practice")
- how many minutes on what
- how I'll know it's done
- leave one day off

For each activity, note what specifically it improves, so I know
why not to skip it. And suggest how I'll check after a month
whether anything's actually changed — some concrete self-test.

This returns a plan you can put on the wall. The last paragraph is worth following through on: progress in a language is slow and hard to see, so without measuring it, people give up before results show. Record yourself talking for three minutes about your day today, and listen back to it in three months — it's the best proof you can give yourself.

Common mistakes

• Letting yourself get corrected mid-conversation. Interruptions kill fluency and lead to silence. Corrections belong after the conversation — and it has to be spelled out explicitly in the instructions, or the model does it on its own.
• Having AI write your essay for you. A submitted piece of text you didn't write teaches you nothing, and the gap shows immediately on a supervised exam where you write alone. Have your own text corrected and explained instead of a new one produced.
• Learning vocabulary as lists. A “word — translation” column gets studied for Friday's quiz, not for speaking. Words belong in sentences and phrases.
• Setting your level too high. A conversation where you understand half of it isn't demanding practice, it's a waste of time. Raise the level gradually, only once the current one feels comfortable.
• Training listening only on AI's voice. Synthetic speech is slower and cleaner than real people. Real recordings are the only practice that matches reality.
• Trusting a model's pronunciation judgment. It's judging a transcript, not a sound. For pronunciation, use native-speaker recordings and shadowing.

The best tools

• Claude in voice mode (mobile) — conversations while walking, roleplay, oral-exam practice. Your level and correction rules can be saved to a Project so you don't repeat them.
• Projects (persistent context) — one place with your level, goal, and rules written down, so every new conversation starts correctly configured.
• Anki or another flashcard app — spaced-repetition words and phrases, ideally pulled from your own mistakes and conversations; the guide is at flashcards with AI.
• A dictionary with pronunciation and examples — for checking what the model claims about typical collocations, and a source of correct pronunciation.
• Podcasts and shows in the target language — the only listening material that talks like real people.
• Your textbook and class notes — still worth it: check disputed grammar rules against them, not against a chat.

What you get out of it

• Time: fifteen minutes a day instead of one hour of tutoring a week means several times more minutes of actual speaking per month — and speaking is what most students are stuck on.
• Money: a large chunk of what a tutor does between lessons — correcting texts, drilling, conversation — you can handle yourself. A live teacher can then focus on what a machine can't do.
• Calm: mistakes in front of AI aren't embarrassing. For a lot of people, that's exactly what finally gets them talking.
• Quality: feedback arrives instantly and specifically — instead of a red squiggly line in a notebook, you know exactly what's wrong and why.

Pro tip

Once a month, hold a conversation one level above where you normally speak, and let the model use more complex language. It'll be uncomfortable and you'll understand less — and that's exactly why it works: when you drop back to your regular level afterward, it suddenly feels easy. Same principle as running uphill before a flat stretch.

And a rule that governs the whole guide: a foreign-language text in your notebook that isn't yours is a loan with high interest. It comes due on the exam, where you write alone with no help. Everything you have AI explain, correct, and walk you through is yours; everything you have it write for you isn't. The honest workflow for a thesis covers how the same principle applies to bigger school projects.