←Back/Product/Codex
IntermediateProduct

Create an Immersive AI Language Tutor

Build a voice-first language practice app that forces immersion by refusing to switch to the user's native language. The AI tutor provides corrections in real-time, within the flow of the conversation, to simulate an authentic learning environment.

Illustration for Create an Immersive AI Language Tutor
Source episode · 00:17:12William Imoh Built a Norwegian Tutor That Refuses to Speak English

with Claire Vo

I said, 'Jeg er på en møte,' that's the wrong tense. So it says, 'It's called et møte, not en møte.' So it corrected me as I said that and then asked, 'What's happening in the meeting?' Then I said, 'Can you speak English?' Uh, and it repeatedly said, 'I'm only gonna speak Norwegian.'

Read the full episodeWatch on YouTube

Before you start

What you need

  • Access to a real-time voice AI model API.
  • A system prompt defining the tutor's persona and rules.
  • A target language for tutoring.

What you’ll make

A functional voice application where an AI tutor converses in a target language, provides in-context corrections, and refuses to switch to English.

Tools used

  • Codex

    OpenAI's cloud-based AI software engineering agent that can execute code, run tests, and handle complex multi-file tasks autonomously.

    VisitCodex

The workflow

Follow the sequence once, then adapt the prompts, checks, and handoffs to your own setup.

4 steps

Step 01

Define the Core Immersion Rule in a System Prompt

The most important feature is the tutor's refusal to break immersion. This behavior must be explicitly defined in the instructions for the language model.

Example prompt
You are a [target language, e.g., Norwegian] language tutor. Your goal is to create an immersive experience. You must only speak [target language]. If the user makes a grammatical mistake, correct it gently and immediately continue the conversation in [target language]. If the user asks you to speak English, you must politely refuse and continue the conversation in [target language].

Make the instruction clear and test it; a prompt alone cannot guarantee consistent behavior.

Step 02

Implement Real-Time Voice Interaction

Use a real-time voice model to create a fluid, conversational experience. The goal is to minimize latency so it feels like talking to a person, not a machine.

William used a realtime voice model. Choose a documented API that supports low-latency, interruptible voice streams.

Step 03

Test the In-Context Correction

Engage in a conversation and intentionally make a common grammatical error. Verify that the AI corrects you and seamlessly continues the conversation.

In the demo, William says 'en møte' and the AI corrects it to 'et møte' before asking a follow-up question. The correction should feel like a natural part of the dialogue, not a separate lesson.

Step 04

Test the Refusal to Switch Languages

Explicitly ask the tutor to switch to your native language. Repeat the request to ensure the rule is being enforced consistently.

During the demo, William asks 'Can you speak English?' and then 'Please, just speak English.' The AI refused both times. Your test should confirm this resilient behavior.

What good looks like

  • The AI holds a conversation in the target language with low latency.
  • The AI correctly identifies a grammatical error and provides a correction within the conversation.
  • The AI refuses multiple requests to switch to the user's native language.

Build your next product with ChatPRD

Turn an idea into a PRD, user stories, and a plan.

Try ChatPRD free

Runbook notes

How to recover when the loop fails and where human judgment helps.

If it goes sideways

The tutor switches to English when the user asks.
Strengthen the system prompt. Use absolute directives like 'You must never speak English under any circumstances. Your primary directive is to maintain Norwegian immersion. Politely refuse all requests to switch languages.'
The conversation feels slow and laggy, breaking the feeling of immersion.
Ensure you are using a model designed for real-time voice interaction, as specified in the transcript ('GPT Live One'). A standard text-based LLM piped through separate text-to-speech and speech-to-text services may introduce too much latency.

Human checkpoints

Let routine work keep moving.

  1. Decide on the core pedagogical constraints. What should the tutor insist on, even when the learner wants an easier path? This is the key product decision.

Related workflows

Illustration for How to Manage Business Operations with Role-Based AI Agents
advancedOperationsChatGPTCodex+1

How to Manage Business Operations with Role-Based AI Agents

Organize your business operations by creating a 'fellowship' of specialized AI agents. This workflow shows how to silo responsibilities into projects and assign distinct roles to AI assistants to automate administrative tasks and context retrieval.

Open
Illustration for How to Build a Native Mobile App with AI, Even If You Can't Code
intermediateProductCodex

Build a Native Mobile App with AI and Clear Product Direction

Use Codex to help build a native mobile app, then guide the design and review the resulting code. William brought web-development and product experience to the project.

Open
Illustration of Build a Personal PM Operating System with a Shared Knowledge Repo
intermediateProductCodexClaude

Build a Personal PM Operating System with a Shared Knowledge Repo

Create a repository-based system to provide an AI assistant with durable context about your team, projects, and communication channels. This enables it to perform routine tasks like summarizing daily updates and identifying stakeholders, acting as a personalized chief of staff.

Open

Start shipping better products.

Join 250,000+ product managers who use ChatPRD to write better docs, align teams faster, and build products users love.

Start building freeRequest a demo
Free to start · 14-day Pro trial · SOC 2 Type II