ChatGPT Prompts to Fix L2 Grammar and Speaking Mistakes: A Teacher’s Toolkit

Your L2 students are already using ChatGPT. The only question is whether they are using it well. Most learners paste a sentence, get a polished rewrite, copy it back, and learn nothing. As ESL teachers, we can change that by giving students a small set of correction-focused prompts that turn the chatbot from a homework shortcut into a patient, judgment-free error coach.

This guide walks through the exact prompts I give my B1–C1 students for fixing grammar and speaking mistakes, the classroom workflow that makes them stick, and the traps that quietly undo the learning.

Why ChatGPT Works for L2 Error Correction

Three things make a large language model uniquely suited to L2 correction: availability, patience, and the ability to explain a rule three different ways without sighing. A class of thirty can’t get individual error analysis on every essay. ChatGPT can. The catch is that it defaults to rewriting, not teaching — so the prompt has to override that habit.

Used well, it gives learners what Stephen Krashen called comprehensible input plus what Merrill Swain called pushed output: feedback on the gap between what they meant and what they actually said. Used badly, it gives them a polished paragraph they didn’t write and can’t reproduce.

The Foundation Prompt: Set the Role Before Anything Else

Before students paste any English, they should set ChatGPT’s role for the conversation. This is the prompt I have my B1+ students save in their notes app and paste at the start of every session:

You are my English error-correction coach. I am a B1 learner. When I send you English, do not rewrite it into perfect English. Instead: (1) identify each grammar, word choice, or word order mistake; (2) explain the rule in simple English; (3) give me one corrected version; (4) ask me a follow-up question using the same grammar so I can practice it. Keep your replies short.

That single prompt does most of the heavy lifting. It forces transparent error tagging instead of silent rewriting, sets a level so explanations don’t drown the student, and ends every reply with productive practice. Adjust the level (A2, B2, C1) to match the learner.

Prompts for Fixing Written Grammar Mistakes

Once the role is set, students can use these targeted follow-ups depending on what they want fixed.

1. The error-tag prompt

“Here is a paragraph I wrote. Mark each mistake with a tag like [TENSE], [ARTICLE], [PREPOSITION], [WORD ORDER], [COLLOCATION], or [SPELLING]. Then list the corrections in a table.”

Tagging is the single biggest upgrade you can make to student-side correction. Instead of seeing a sea of red ink, the learner sees patterns. Three [ARTICLE] tags in one paragraph tell them exactly what to study tonight.

2. The minimal-change prompt

“Correct only the grammar errors. Do not change my vocabulary, style, or sentence length. Keep my voice.”

This stops ChatGPT from quietly upgrading a learner’s B1 essay into C2 prose. The student needs to see their sentence, fixed — not a stranger’s sentence pasted in.

3. The explain-don’t-fix prompt

“Don’t give me the answer yet. Tell me what kind of mistake is in this sentence, and ask me a question to help me find it myself.”

This is the Socratic-tutor mode. Brilliant for intermediate learners who already know most rules but need to retrieve them under pressure.

4. The contrastive prompt

“Show me my sentence and the corrected sentence side by side. Then explain what changed and why.”

The side-by-side forces attention on form. Without it, learners read only the corrected version and never internalize the gap.

Prompts for Fixing Speaking Mistakes

Speaking correction is trickier because ChatGPT can’t hear the learner directly. The workaround is the phone’s built-in voice-to-text, or the ChatGPT app’s voice mode, which transcribes the student’s speech into text the model can then analyze.

The speak-and-transcribe loop

Have students speak a 60-second response into voice-to-text, then paste the raw transcript with this prompt:

This is a transcript of me speaking English. Treat it as spoken English — don’t punish me for fillers like “um” or “you know”. Identify: (1) real grammar mistakes; (2) unnatural word choices a native speaker wouldn’t use; (3) one collocation I should learn. Ignore filler words. Reply in a list.

The “treat it as spoken English” line is critical. Without it, the model will flag every false start as a mistake and demoralize the student. Spoken language is messier than written language; the correction should respect that.

The fluency-first prompt

“Give me three sentences I could have said more naturally, and tell me one phrase I should memorize from this topic.”

This shifts focus from accuracy to naturalness, which is where most B2+ learners get stuck. They are grammatically correct but stilted. “How would a native speaker say this?” is a more useful question for them than “What did I get wrong?”

The pronunciation-by-spelling trick

Voice-to-text reveals pronunciation errors indirectly: if the student says “I sink so” and the transcript shows “I sink so”, they have a /θ/ vs /s/ problem. Have them ask:

The transcript says “sink” but I meant “think”. What sound am I mispronouncing? Give me three minimal pairs to practice tonight.

That’s a diagnostic the student can run on themselves, every day, for free. It’s not a perfect pronunciation coach, but it catches the high-frequency consonant substitutions that the L1 makes predictable.

Classroom Workflow: A 4-Step Cycle

Handing students a list of prompts is not enough. They need a routine. Here is the cycle I run in a 50-minute lesson, adaptable to homework:

1. Produce — Student writes a paragraph or records a 60-second answer to a prompt (e.g. “Describe your weekend.”). No AI yet.
2. Self-edit — Student rereads or rereads-the-transcript and circles three things they think are wrong. This metacognitive step is what makes the AI feedback stick later.
3. Consult — Student opens ChatGPT, pastes the role prompt, then submits their work with the error-tag prompt.
4. Reproduce — Student rewrites the paragraph from scratch, without looking at the AI version, then compares. The gap between attempt 1 and attempt 2 is the actual learning.

The reproduce step is non-negotiable. Without it, the student has read a correction but not produced one. Reading correct English is not the same skill as writing it.

Common Pitfalls (And How to Coach Around Them)

Pitfall 1: The polish trap

Students submit homework, get a polished AI rewrite, hand it in. Catch this in two ways: ask for handwritten drafts, and assign the reproduce step in class without a phone.

Pitfall 2: The over-correction spiral

ChatGPT will happily flag stylistic preferences as “mistakes”. A B1 learner sees 14 red tags on a paragraph and gives up. The level-specific role prompt (“I am a B1 learner”) helps. So does telling students: “Only fix the grammar tags this week. Ignore the style notes.”

Pitfall 3: Hallucinated rules

The model occasionally invents a grammar rule that doesn’t exist or misapplies one. Teach students one verification habit: “If the rule sounds weird, ask me in class tomorrow.” The teacher remains the final authority. AI is the volume-multiplier, not the replacement.

Pitfall 4: Passive consumption

Many learners scroll the AI reply, nod, and close the app. The follow-up question built into the role prompt forces output, which is where retention lives.

A Sample Correction Dialogue (B1 Student)

To make this concrete, here is a real-style exchange a student might produce after pasting the role prompt:

Student: “Yesterday I go to the market and buy three vegetable for cooking dinner with my mother.”

ChatGPT (in coach mode):

• [TENSE] “go” → “went” (past simple, because of “yesterday”)
• [TENSE] “buy” → “bought”
• [PLURAL] “three vegetable” → “three vegetables” (countable plural after a number)

Corrected: “Yesterday I went to the market and bought three vegetables to cook dinner with my mother.”

Your turn: Tell me about something you did last weekend. Use at least three past simple verbs.

Notice the model didn’t simply rewrite the sentence — it tagged, explained, corrected, and prompted output. That’s what the role prompt is buying us.

Adapting the System for Different Levels

A2 learners

Limit error tags to three categories: TENSE, ARTICLE, PLURAL. Anything more is cognitive overload. Add the line “Use only words from the most common 1000 English words when you reply” to the role prompt.

B2 learners

Add COLLOCATION and REGISTER to the tag list. This is the level where “do a mistake” vs “make a mistake” becomes the bottleneck, and where formal-vs-informal choices start to matter.

C1 learners

Switch from error correction to nuance feedback. Prompt: “My grammar is fine. Tell me which sentences sound non-native and rewrite them in three different registers: academic, journalistic, casual.”

What Teachers Should Still Do (That AI Can’t)

It’s worth saying plainly: ChatGPT does not replace the teacher. It replaces the worksheet. Three things still need a human in the room:

• Diagnosis of patterns over time — only the teacher sees the student’s whole error history across weeks. AI sees one conversation.
• Motivation and accountability — a chatbot cannot notice that the student looks defeated, or hasn’t done homework in three days.
• Real-time pronunciation modeling — voice-to-text gets you part of the way. Lip shape, mouth position, and the actual th sound need a teacher’s mouth on camera.

Frame it for students this way: ChatGPT is the gym they go to alone six days a week. You are the coach they see once a week who tells them what to train next.

A Mini Lesson Plan You Can Run Tomorrow

If you want to introduce this system in a single 50-minute lesson, try this structure:

1. (5 min) Hook — Show students an AI rewrite of a B1 paragraph. Ask: “What did the learner actually learn from this?” Get them to articulate the polish trap themselves.
2. (10 min) Role prompt — Have students copy the foundation prompt into their notes app. Discuss why each line is there.
3. (10 min) Produce — Students write a short paragraph about their last holiday on paper. No phones.
4. (15 min) Consult and tag — Students paste the role prompt + their paragraph into ChatGPT, get tagged feedback, and copy the tag categories into a notebook page titled “My Mistake Patterns”.
5. (10 min) Reproduce — Students close the AI and rewrite the paragraph from scratch. Pair-share to compare.

End the lesson by setting the homework: run the same cycle on a different topic at home, screenshot the tag list, and bring it next class. Over a month, those screenshots become a personal error profile that no textbook could ever build.

Final Thought

The teachers who win the next five years won’t be the ones who ban AI in their classrooms. They’ll be the ones who teach students how to make AI teach them. A B1 student with the right correction prompt is, for the first time in the history of language learning, getting individualized feedback on every sentence they produce. That’s a quiet revolution. Our job is just to make sure they’re asking the chatbot good questions.

Kaynaklar

• Cambridge Üniversitesi Yayınları — research and frameworks on second language acquisition and feedback.
• İngiliz Konseyi — CEFR level descriptors and teaching resources for ESL practitioners.
• TeachingEnglish (British Council & BBC) — practical methodology articles on error correction and feedback.
• OpenAI — official documentation for ChatGPT prompt design.

More Worksheets

Typhoon Season in Taiwan | Elementary ESL Reading Worksheet PDFThe 7 Most Common AI Mistakes in ESL Materials (and How to Fix Them)How to Fact-Check AI-Generated Content Before Your Next ESL Class