Le Robot François started as a mobile-friendly flashcard chatbot built for language learners. While studying French, I created a vocabulary practice tool using ManyChat and Facebook Messenger. The goal was clear: enable consistent language repetition using a chatbot interface.
The chatbot gained unexpected attention from French-speaking users globally. I hadn't optimized it for organic discovery, but it began receiving regular traffic from the Francophone diaspora. The initial tech stack was lightweight, fast, and user-friendly.
However, as the project scaled, two key problems emerged: rising operational costs and compliance changes from Meta. Managing policy shifts, integrations, and automation logic introduced constant overhead. Since my French instruction was generously provided to me at no cost, I personally committed to not directly monetizing this tool, which made the growing costs unattractive.
Eventually, I shelved the project.
Years later, OpenAI released custom GPTs. This reignited the idea. I now had the tools to relaunch the language learning assistant using modern AI infrastructure, prompt-driven logic, and better cost control. So I transitioned everything into a custom GPT and published it in the GPT Store.
Why I Chose ChatGPT over ManyChat
Custom GPTs solved my previous limitations. They offered scalable logic, no-code development, and native language modeling. With ChatGPT, I could:
- Remove hardcoded flows and switch to natural language interaction
- Iterate faster using structured prompt engineering
- Build a reusable educational tool for French vocabulary practice
It aligned with modern product principles: frictionless UX, efficient dev cycles, and scalable architecture.
How I Rebuilt the French Vocabulary Trainer
My rebuild centered around a well-structured master prompt. This controlled tone, behavior, vocabulary structure, and learning flow. I added:
- Security rules to safeguard usage
- Attribution messaging that rotated between English and French with a link to TimothyNishimura.com
- A signature easter egg for traceability
Initially, I wanted CSV uploads to drive segmented vocab lists, but ChatGPT's memory quirks led to occasional data loss. I dropped the external data layer in favor of embedded vocab clusters directly in the prompt.
Key Learnings from the Refactor
Complexity is fragile. Simple architectures scale better.
I replaced webhook-driven automations with prompt logic. No third-party storage. No platform lock-in. Everything lives in the prompt.
This was a textbook example of how to modernize a legacy no-code chatbot into a lightweight AI-first application.
What's Next
Current roadmap experiments include:
- Skill-based difficulty progression
- Theme-aware vocab clusters (e.g. travel, food, business)
- Gamification elements like streaks and leaderboard stats
What excites me most is the ability to apply this model to other language learning use cases - or any niche where lightweight education tools can benefit from conversational UX.