Protégé

• 

  HomeScreen

Inspiration

I've studied for hours, felt completely ready, and then blanked when someone asked me to explain it.

Through my study of psychology, I learned that this is called the illusion of explanatory depth. We think we understand something until we have to explain it out loud. Re-reading and highlighting feel like studying, but they never test that.

I'm also building a learning app called EdU, so I think a lot about how people actually learn. The best way to find out what you don't know is to try to teach it.

What it does

Protégé lets you teach a topic to an AI that plays a confused student.

• You pick a topic and explain it in your own words.
• The AI asks the kind of question a real learner would ask. It never explains the topic back to you.
• It highlights the exact sentences that confused it and tells you why.
• A confusion meter shows how lost it is. You fix your explanation, try again, and watch the number drop.
• At the end, a summary shows what you explained well and what you missed.

How I built it

• Frontend: React and TypeScript
• Backend: FastAPI (Python)
• AI: Claude, through an API. The key stays on the server, never in the browser.
• Your explanation is split into numbered sentences before it goes to the AI. The AI replies with structured data: a follow-up question, a confusion score, scores for accuracy, clarity, missing steps, and jargon, and a list of which sentence numbers confused it. The frontend uses that list to highlight the right sentences.
• The prompt tells the AI to act like a curious student who knows nothing yet, asks one question at a time, and doesn't flatter.

I built it quickly with AI coding tools. The idea, the design choices, and the learning science are mine.

Challenges I ran into

• Keeping the AI confused. AI wants to be helpful, so it kept filling in the gaps and explaining the topic back to me. I had to tighten the prompt so it only reacts to what the student actually said. [Add anything you changed or noticed while fixing this.]
• Making the highlighting match. Splitting the text into numbered sentences first was what made the highlights land on the right lines.
• Time. I built this in one night, so I cut everything except the core loop: teach, get confused, fix, repeat.
• Teach-back challenge: the AI retells your explanation using only what you taught it, mistakes included, and you tap the sentences that are wrong to correct them.

What I learned

Teaching shows you gaps that studying hides. On the technical side, I learned that getting reliable structured output from an AI depends as much on the prompt as on the code, and that a small tool that works every time beats a big one that doesn't.

What's next for Protégé

• A classroom mode where teachers can see which concepts students struggle to explain.
• Voice input, so you can talk through a topic instead of typing.
• Connecting it to EdU as a "teach it back" mode.

Built With

• anthropic-ai
• claude
• css
• emergent
• fastapi
• html5
• javascript
• python
• react
• typescript