How to Build Custom English Learning Prompts for AI Tutors: A Complete Guide

Create a persistent "English Coach" system prompt that defines your level, goals, and session structure, then chain specialized AI tools like Gemini Live for speaking, Canvas for writing, and quiz generators for review to close the input-output-feedback loop.

The byoungd/English-level-up-tips repository provides a production-ready framework for converting generic AI assistants into personalized English tutors. According to the source code in docs/threads/part-1/7-ai.md, the key lies in engineering custom Gems (programmable system instructions) that persist your learning profile across sessions and orchestrate multi-tool workflows covering listening, speaking, reading, and writing.

The Architecture of an AI English Tutor

Effective English learning with AI relies on closing the input-output-feedback-review loop. The repository outlines a four-stage architecture where custom prompts drive each stage.

The Four-Stage Learning Cycle

1. Input: Expose yourself to authentic material (articles, podcasts, videos).
2. Output: Produce language through speaking or writing tasks.
3. Feedback: Receive targeted corrections on high-frequency errors.
4. Review: Convert mistakes into active-recall quizzes and spaced-repetition flashcards.

This cycle is implemented through specialized components configured via specific prompt instructions found in the AI chapter.

Core Components and Source References

• Gem (Custom English Coach): A persistent system prompt that stores your curriculum, proficiency level, and error history. Defined in lines 58-73 of docs/threads/part-1/7-ai.md.
• Gemini Live: Real-time spoken practice module that forces verbal output and delivers instant corrective feedback. Configuration template at lines 100-107.
• Guided Learning: Step-by-step reading comprehension that requires prediction and reflection rather than simple translation. Usage pattern described at lines 57-60.
• Canvas: Persistent document editing environment for writing practice with annotated revision history. Workflow detailed at lines 97-100.
• Quiz/Flashcard Generator: Automated conversion of input materials into active-recall assets. Prompt templates at lines 27-34.

Step-by-Step: Building Your English Coach Gem

The foundation of your AI tutoring system is a static instruction block stored in Gemini's custom Gem. This prompt persists across all conversations, ensuring the AI maintains context of your goals and recurring mistakes.

According to lines 58-73 of docs/threads/part-1/7-ai.md, use this template:

You are my English Level‑Up Coach.  
I am a Chinese speaker at B1‑B2 level, aiming to boost oral, listening and work‑scene English in 12 weeks.  
Your tasks:
1️⃣ Speak English by default, but explain complex grammar in brief Chinese on request.  
2️⃣ Each session = 20‑30 min and must contain: warm‑up, input material, output task, error correction, debrief.  
3️⃣ Prioritise high‑frequency, transferable errors; avoid nit‑picking.  
4️⃣ Record my recurring mistakes and schedule weekly revision.  
5️⃣ When I say “Start today’s lesson”, launch the next module.  
6️⃣ If I upload an article, résumé, or video subtitle, turn it into a lesson plan.  
7️⃣ End each session with: key expressions learned, top errors, homework, next‑session focus.

Store this in the Gem's instructions field. Once configured, saying "Start today's lesson" triggers the AI to load your day's sub-goal (warm-up → input → output → feedback → review).

Specialized Prompt Templates for Each Skill

After initializing your Gem, invoke specific prompt patterns for different learning modalities. These templates are derived directly from the repository's AI chapter.

Speaking Practice with Gemini Live

For real-time spoken practice, use this concise system prompt when starting a Live session (lines 100-103):

Please act as my speaking coach. We will have a natural English conversation for 15 minutes. Keep your turns short. Interrupt me when necessary only if my sentence is hard to understand. After every 3 rounds, give me brief feedback on grammar, word choice, and pronunciation priorities.

This forces output production under time pressure while receiving immediate corrective feedback.

Reading Comprehension and Guided Learning

When processing articles or transcripts, avoid simple translation. Instead, use Guided Learning prompts that force active prediction (lines 27-30):

Create a quiz about this material. Start with 5 easy comprehension questions, then 5 harder inference questions. After each answer, do not tell me only whether it is right or wrong. Explain why.

This targets deeper processing and retention compared to passive reading.

Vocabulary and Flashcard Generation

To build spaced-repetition assets from input material, use the flashcard-focused prompt (lines 31-34):

Create flashcards about this material. Focus on high-frequency vocabulary, collocations, and sentence patterns that are useful in real conversations, not just rare difficult words.

This filters for transferable high-frequency expressions rather than obscure vocabulary.

Writing Feedback with Canvas

For writing practice, the Canvas workflow emphasizes self-correction before model answers (lines 97-100):

Here is my draft. Do not rewrite everything immediately. First identify the most important mistakes and weak sentences. Explain why they are weak. Then ask me to revise them myself. After I revise them, show me a stronger version for comparison.

This method builds metalinguistic awareness by requiring you to revise errors yourself before seeing polished alternatives.

Multi-Tool Orchestration Strategies

Different AI tools excel at specific tasks. As noted in section 9 (lines 1-5) of the AI chapter, delegate specialized work to optimize results:

• ChatGPT Study Mode: Deep grammar explanations and nuance comparisons (e.g., distinguishing "effective" vs. "efficient" vs. "practical").
• Claude Projects: Long-form reading analysis and complex document processing.
• Perplexity Spaces: Material discovery and source verification for authentic input.
• DeepL Write: Final polishing and style refinement of completed drafts.

Cross-tool workflow example: Use Perplexity to find a technical article, feed it to your Gem for Guided Learning, practice discussing it in Gemini Live, then polish your written summary in DeepL Write.

Summary

• Engineer a persistent Gem using the system prompt template from lines 58-73 of docs/threads/part-1/7-ai.md to maintain long-term learning context.
• Close the learning loop by combining Input (Guided Learning), Output (Gemini Live), Feedback (Canvas), and Review (quizzes/flashcards).
• Use specific prompt structures that prioritize high-frequency errors over nit-picking and require active self-revision before providing model answers.
• Orchestrate multiple AI tools by assigning ChatGPT for grammar deep-dives, Claude for reading, Perplexity for discovery, and DeepL Write for polishing.

Frequently Asked Questions

How do I prevent the AI from over-correcting minor grammar mistakes?

Configure your Gem instructions to explicitly prioritize high-frequency, transferable errors while avoiding nit-picking. The prompt template at lines 58-73 includes the directive: "Prioritise high-frequency, transferable errors; avoid nit-picking." This constraints the AI to focus only on mistakes that significantly impede communication or reflect systemic gaps.

Can I use these prompts with ChatGPT if I don't have access to Gemini?

Yes. While the repository emphasizes Gemini's Gems and Live features for persistence and real-time speech, the underlying prompt architectures work across platforms. Store the English Coach system prompt in ChatGPT's Custom Instructions or Claude's Projects feature. However, you will need external voice tools for real-time speaking practice since ChatGPT lacks Gemini Live's native interruption and pronunciation feedback capabilities.

How do I ensure the AI remembers my mistakes across sessions?

The Gem (or equivalent persistent memory feature) stores your recurring mistake log in its instructions field. Include the directive: "Record my recurring mistakes and schedule weekly revision." When you upload new material or start a session, reference this log by asking: "Review my recurring errors from last week before we start." This maintains continuity without manual tracking.

What is the optimal session length when using these AI prompts?

The repository recommends 20-30 minute sessions structured as: warm-up, input material, output task, error correction, and debrief. This duration balances intensity with retention while preventing cognitive overload. The Gem instructions explicitly define this time constraint to keep the AI pacing appropriately.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}