How to Build an Interactive 3D Website With GPT-6 Astra

We gave ChatGPT 6 Astra one plain-English prompt, no images, no architectural plans, no 3D model and no code. It researched Pune's Shaniwar Wada and built an interactive browser experience. Exact prompt included.

Most AI demos end with an answer, an image or a block of code you still need to assemble yourself.

This one ended with a working website.

We asked ChatGPT 6 Astra to create an interactive 3D explainer of Shaniwar Wada, Pune's iconic 18th-century Peshwa fortification. We did not upload a single reference image. We only supplied the coordinates and gave it permission to research the building.

From that, Astra found visual references, studied the surviving architecture, built the 3D scene, added interactive viewpoints and deployed the experience to a live URL.

One prompt produced the functional prototype. A few short art-direction notes refined the colours, landscaping and interface before publishing.

The Exact Prompt

Hey astra, this is my first ever message to you. Get me an interactive explainer
3d build of Pune Shaniwar Wada. i don't have pics with me now. just putting
coords here: 18.51927° N, 73.85424° E, use this or do a search to get the best
possible images of the iconic building. Whatever you feel. Thank you

That was it.

No framework. No design system. No list of interactions. No instructions about Three.js, WebGL, materials, lighting or deployment.

The prompt gave Astra four useful things:

• •A clear output — an interactive 3D explainer
• •A specific subject — Shaniwar Wada
• •A verified location — exact coordinates
• •Permission to research — search for the best available references

The model handled the implementation choices itself.

1Research the Building Before Modelling It

Shaniwar Wada is a difficult subject because the building people imagine and the site that survives today are not the same thing.

The original palace complex was largely destroyed by fire in 1828. What remains includes the fortified walls, bastions, gates, foundations, gardens and water structures. A responsible reconstruction has to distinguish visible evidence from historical interpretation.

Astra searched for photographs of:

• •Delhi Darwaza, the main north-facing entrance
• •The paired cylindrical bastions
• •The studded wooden gate
• •The carved timber gallery above the entrance
• •The internal lawns, pathways and palace foundations
• •Hazari Karanje, the lotus-shaped fountain

This mattered more than raw polygon count. A detailed model built from the wrong architectural assumptions would still be wrong.

2Build the Entire Scene in the Browser

The experience was generated as a browser-native 3D scene. The fort, trees, lawns, pathways, foundations, doors, railings and people are all rendered interactively rather than flattened into a video.

That gives the viewer control:

• •Drag to orbit around the structure
• •Scroll or pinch to zoom
• •Jump between four composed viewpoints
• •Switch between daylight and evening light
• •Turn on automatic orbit

The four viewpoints were designed for different kinds of understanding:

The result is closer to a small digital museum exhibit than a conventional webpage.

3Art-Direct the Model With Plain Language

The first working version proved the interaction, but its visual treatment was weak. The stone was too pale, the masonry texture felt artificial, the landscaping was sparse and a geometry bug stretched parts of the wall into horizontal slabs.

The correction did not require editing code manually. We gave Astra direct visual feedback:

Wrong colors, texture, no landscaping. Not aesthetic at all. You put some
horizontal lines at wall height. Do better.

We also supplied a screenshot of another architectural 3D study as an art-direction reference.

Astra identified the wall-generation bug, reworked the materials using the real gate photographs, corrected the roof and timber gallery, and rebuilt the grounds with:

• •Weathered brown basalt
• •Warm carved timber
• •Stone paving and entrance steps
• •Formal lawns and clipped shrubs
• •Mature perimeter trees
• •Garden terraces and pathways
• •Small human figures for scale
• •Softer daylight on an ivory presentation field

This is the useful shift: feedback can now describe the visual problem rather than prescribe the technical fix.

4Strip the Interface Back

The second version looked much better, but it carried too much editorial framing: a header, coordinates, architectural-study labels and supporting copy.

The final interface was reduced to what the experience actually needed:

• •The name: Shaniwar Wada
• •Four viewpoint controls
• •A short description that updates with each view
• •Evening and auto-orbit controls

The selected description was moved to the top-right corner so the building remained the centre of the composition.

This is a good rule for AI-generated interfaces: the model will often add explanatory furniture because it wants to prove that it understood the brief. Remove anything the user would not miss.

5Export and Publish the Experience

Once the interactive build was approved, the final instruction was simple:

Possible to quickly export it as a site? So I can check it in browser full screen.

Astra packaged the experience, published it as a private site and returned a browser-ready URL. The full-screen version preserved the camera controls, responsive layout and live 3D rendering.

What Astra Actually Did

What This Used to Require

Until very recently, an experience like this meant stitching together a full production pipeline.

You would research the site, collect references, model the architecture in Blender or SketchUp, create materials and lighting, export and optimise the assets, build the interaction in Three.js, design the interface, test it across devices and then configure hosting.

That usually meant several specialist tools, multiple handoffs and days or weeks of work before anyone could open a link.

With Astra, the same path became a conversation. Describe the experience, review the result, give visual feedback and publish. Blender, texture software and hand-written WebGL are no longer prerequisites for getting a convincing interactive prototype in front of people.

For conservation-grade accuracy, measured drawings and expert review still matter. But for concept development, storytelling and public-facing prototypes, the barrier to entry has changed completely.

Where You Can Use This

• •Heritage and tourism — turn monuments, forts and historic districts into interactive visitor experiences
• •Museums and education — let learners explore objects, buildings and historical reconstructions instead of only reading about them
• •Architecture and real estate — create early walkthroughs for unbuilt spaces, masterplans and property launches
• •Products and engineering — build explorable product demos, exploded views and feature explainers
• •Events and exhibitions — preview stages, stalls, galleries and spatial concepts before production
• •Brand campaigns — replace a passive landing page with a small world people can move through
• •Urban planning — communicate streetscape, landscape and public-space proposals in a form non-specialists can understand

How to Prompt Astra for a Similar 3D Build

Use this structure:

Create an interactive 3D explainer of [specific building, object or place].

Reference information:
- Location or coordinates: [coordinates]
- Important features: [features, if known]
- Use web research to find accurate visual references.

The experience should let users [orbit / zoom / switch views / inspect parts].
Keep the interface minimal. Clearly label any geometry that is interpretive
rather than based on measured plans.

When complete, export it as a responsive full-screen website.

You can leave the technical stack unspecified unless you have a hard constraint. Describe the subject, the experience and the standard of accuracy you expect. Let the agent choose the implementation.

The Real Takeaway

AI is moving beyond generating assets for a workflow. It can now carry the workflow itself: research the subject, make design decisions, write the code, test the interaction, revise the art direction and publish the result.

The first prompt gets you to something real. Taste gets you from real to good.