An AI English tutor works well when you set it up once and then follow a routine. The setup is a short "tutor brief" that tells the AI your level, your goal, how and when to correct you, and how long a session lasts. The routine is four weeks: measure where you are, build fluency, fix three repeated mistakes, then test yourself against your first recording. Without the brief and the routine, most people chat for a few days, get flattered, and stop.
This page is about running a tutor over weeks. If you want single prompts for speaking, writing and vocabulary, start with learn English with AI and come back here when you want a system.
Your tutor brief (copy, fill in, paste)
Paste this at the start of every session, or save it in your app's custom instructions if it has them:
"You are my English tutor. My level: B1. My goal: speak confidently in meetings at work within two months. My first language is Uzbek; I also know Russian. Correction mode: recast (details below). Correct at most three mistakes per session, and focus on my target errors: articles, present perfect, he/she. Explain in simple English; if I write 'uzbekcha', explain in Uzbek. Session format: 3-minute warm-up, 10-minute task, 5-minute feedback. Ask one question at a time. Never praise me unless something is specifically good, and say what."
Every line in it has a job:
| Setting | What to write | Why it matters |
|---|---|---|
| Level | A2, B1, B2… (take a level test if unsure) | Stops the AI using C1 vocabulary with a beginner |
| Goal | One concrete goal with a date | Keeps topics relevant: meetings, IELTS, travel |
| First language | Uzbek, Russian or both | The AI can predict your typical errors |
| Correction mode | Live, recast, summary or log (see below) | Decides whether you get interrupted |
| Correction limit | Usually 3 per session | Ten corrections at once are forgotten by tomorrow |
| Target errors | Your 2–3 most frequent mistakes | Turns random feedback into a plan |
| Feedback honesty | "Don't praise unless specific" | AI models tend to be too positive |
Four correction modes and when to use each
| Mode | What the AI does | Best for | Tell it |
|---|---|---|---|
| Live | Stops you and corrects the mistake immediately | Drilling one grammar point; pronunciation of a single word | "Stop me every time I make a mistake with articles." |
| Recast | Repeats your idea correctly in its reply and continues the talk, without comment | Fluency practice, nervous speakers | "Don't stop me. Just repeat my sentence correctly in your answer." |
| End-of-session summary | Collects mistakes and lists the top three at the end | Most sessions; exam practice | "At the end, show my three most important mistakes with corrections." |
| Silent log | Keeps a list and shows it only when you ask | Long role-plays and mock interviews | "Keep a list of my mistakes; show it when I type 'log'." |
A recast looks like this. You say "Yesterday I have been to the bank." The tutor answers "Oh, you went to the bank yesterday? Was it busy?" It doesn't break the conversation, but you hear the right form. Recasts are gentle, so combine them with an end-of-session summary, or you may never notice them.
One 20-minute session
- Warm-up (3 minutes). Two easy questions about your day. Speak, don't type, if the app has voice mode.
- Main task (10 minutes). One task from your week's plan: a role-play, a discussion question, or describing a photo from your phone.
- Feedback (5 minutes). The three corrections, each with a new example sentence. Say every corrected sentence aloud twice.
- Recycle (2 minutes). Ask: "Give me two questions that make me use today's corrections." Answer them.
Then write the three corrections into one error log (a notebook or a note on your phone). The log, not the chat history, is your record of progress.
The 4-week routine
Five sessions a week, 20 minutes each: about 7 hours in a month.
| Week | Aim | Correction mode | Tasks |
|---|---|---|---|
| 1: Diagnose | Find your real mistakes | Silent log, summary on day 5 | Day 1: record yourself answering 5 questions for 1 minute each (your baseline). Days 2–5: free conversation on work, study, family, plans. On day 5 ask: "Which three mistakes did I repeat most this week?" These become your target errors. |
| 2: Fluency | Speak longer without stopping | Recast + summary | Answers of 45–60 seconds. Role-plays from your life: a call with a client, a question at the bank, small talk with a new colleague. |
| 3: Accuracy | Fix the three targets | Live, targets only | Short drills: "Ask me 10 questions that force me to use the present perfect." Then a normal conversation where the tutor stops you only for the targets. |
| 4: Test | Measure the change | Summary | Days 1–3: exam-style or real-task simulation (a mock interview, an IELTS-style Part 1). Day 4: repeat the week 1 questions and record. Day 5: compare the two recordings yourself: pauses, length, target errors. |
At the end of week 4, choose three new targets and start again. The comparison of two recordings is the most honest progress check you have, because it uses your own ears, not the AI's opinion.
If you don't know where to find target errors, the list of mistakes Uzbek speakers make is a good menu to choose from.
Where an AI tutor falls short
- It doesn't hear you like a person does. Speech recognition often "fixes" your pronunciation before the AI sees the text, so a W/V or TH problem may never be reported. Check pronunciation with a teacher or a dedicated pronunciation tool.
- It praises too much. Many models are trained to be encouraging. That's why the brief says "don't praise unless specific".
- It drifts. After 20 minutes, the level and correction mode may slip back to default. Paste the brief again or start a new chat.
- It can be confidently wrong. Grammar explanations are usually right for common points and less reliable for rare ones. When a rule surprises you, check it in a learner's dictionary or grammar book; the Uzbek guide on AI mistakes and hallucinations shows typical examples.
- It has no stake in your result. A human tutor notices you've skipped three sessions. An AI doesn't. Put your sessions in your calendar.
Privacy settings to check before week 1
Open the settings of the app you use and look for three things: whether chat history is saved, whether your conversations may be used to train models (some services let you switch this off), and whether the app has a "memory" that keeps facts about you between chats. Decide on each one deliberately.
Whatever the settings, don't put passport or card numbers, passwords, your full address, or confidential work documents into a tutor chat. For role-plays, invent the company name and the numbers. If a child is using an AI tutor, a parent should read the app's age rules and privacy page first.
AI scores are practice numbers
If your tutor says "That answer was Band 6.5" or "You are B2", treat it as a rough signal for comparing your own sessions. Official IELTS bands come only from IELTS examiners, and national certificate levels only from the official testing agency. An AI can give a different number for the same answer if you ask twice. To choose between tools with different features, see English learning apps compared.
Exercise: pick the right setting
Choose the best correction mode (live, recast, summary or silent log) for each learner, then fix the weak brief in item 7.
- 1. Dilshod freezes when he's interrupted and wants to speak more freely.
- 2. Malika keeps saying he about her sister and wants to stop this week.
- 3. Timur is doing a 15-minute mock job interview and wants it to feel real.
- 4. Aziza has a normal 20-minute session and wants three things to work on.
- 5. Bekzod wants to practise the pronunciation of th in ten words.
- 6. Nigora is in week 1 and doesn't yet know her typical mistakes.
- 7. Weak brief: "Be my teacher and correct everything." What is missing?
Answers
- 1. Recast. He hears correct forms without being stopped; add an end-of-session summary so he notices them.
- 2. Live, for one target only. "Stop me every time I use he/she wrongly." Everything else can wait.
- 3. Silent log. Interruptions would ruin the simulation; he reads the log afterwards.
- 4. End-of-session summary with a limit of three corrections.
- 5. Live, and remember the limit: speech recognition may not catch everything, so record yourself and compare with a dictionary's audio.
- 6. Silent log for the week, then a summary of repeated mistakes on day 5.
- 7. Level, goal, first language, correction mode, a limit on corrections, target errors and session length. "Correct everything" gives you a wall of red ink and no plan.
Your tutor, with a voice
BandMate is built around the speaking part of this routine: you talk to Bandi by voice in everyday conversations, work role-plays and an IELTS Speaking mock, and get feedback on grammar, vocabulary and pronunciation with explanations in Uzbek, Russian or English. You can use it for the week 1 baseline and the week 4 comparison. Its scores are estimates for practice, not official results, so keep your own recordings as the real measure. For an exam-focused version of the same idea, the Uzbek guide AI IELTS Speaking mock walks through a full timed session.
Add BandMate as a preferred source on Google
