Look Closer

An AI-powered curiosity engine that notices what catches your eye and gently guides you through the hidden stories inside great masterpieces.

Comment

Inspiration

Most AI learning experiences begin the same way:

You ask a question.

The AI answers.

But when we stand in front of a masterpiece, we often don't know what to ask.

Our attention settles long before we form a question.

That inspired us to explore a different interaction model.

What if AI could notice where our attention naturally settles, and guide curiosity from there?

What it does

Look Closer is an AI-powered curiosity engine for masterpieces.

Instead of waiting for visitors to ask questions, Look Closer quietly observes where their attention naturally settles and gently invites them to explore the hidden stories inside an artwork.

For my first experience, we chose Leonardo da Vinci's The Last Supper.

As visitors move through the painting, GPT-5.6 generates curator-style notes that respond to their exploration. Rather than repeating the same explanation, the experience adapts based on where visitors linger, what they have already discovered, and how they continue exploring.

The result feels less like chatting with an AI and more like walking through an exhibition with a thoughtful museum curator.

How we built it

Look Closer was built using Codex and GPT-5.6.

Tech Stack

  • Next.js
  • TypeScript
  • Tailwind CSS
  • OpenAI Responses API
  • GPT-5.6 Structured Outputs

Attention Engine

The application continuously observes visitor attention through cursor movement, hover duration, revisit behavior, and exploration history.

Invisible hotspots provide stable spatial context while GPT-5.6 generates contextual curator notes based on each visitor's journey.

Codex

Codex accelerated development throughout the project by helping scaffold the application, implement the interaction system, build reusable components, iterate on the UI, and rapidly refine the overall experience.

GPT-5.6

GPT-5.6 powers the AI Curator.

Rather than returning static descriptions, GPT-5.6 receives structured information about the visitor's attention, previously explored hotspots, interaction stage, and artwork context.

It then generates contextual curator notes using Structured Outputs, allowing every exploration to feel slightly different while remaining reliable and consistent.

Challenges

The biggest challenge wasn't generating stories.

It was designing how AI should behave.

Most educational AI systems react after users ask questions.

I wanted AI to behave more like an experienced museum curator—quietly observing, recognizing attention, and speaking only when it genuinely adds value.

Designing that interaction became the core challenge of the project.

Balancing immersion, stability, and dynamic AI generation required combining deterministic hotspot data with GPT-5.6 generated curator notes so the experience remained both reliable and personal.

What's next

The Last Supper is only our first demonstration.

The same interaction model can naturally extend to:

  • Museums
  • Historical photography
  • Scientific illustrations
  • Architecture
  • Maps
  • Cultural heritage
  • Comics
  • Educational collections

I believe attention can become a new interface for learning.

Instead of asking users to search for knowledge, AI can recognize what naturally draws their attention—and help them discover something meaningful from there.

Built With

Share this project:

Updates