A method, not a manual

Investigate the history of your own home

This is a rough map of how we built Beneath Evering Road, a page that starts at a London front door and ends 410 million years down, on a buried mountain range. None of it is specific to that house. Yours will work too.

1  Start from something real

Take a photo of the front of your home, or grab a street-level capture. This one image is the anchor for everything. Every generated frame will descend from it, which is what keeps the whole thing honest instead of drifting into a generic "olden days house."

2  Let a coding agent do the digging

Point an AI coding agent (we used Claude Code) at the public record and let it chase citations. For a UK address that means census returns, old street directories, historic maps (Booth's poverty survey, Rocque's 1746 map, Layers of London, the Victoria County History, Domesday), local museum and archive collections, then geological survey data for what the ground itself was before anyone thought to live on it. Your country will have equivalents.

The agent's job is to assemble a dated timeline of "who and what was here," with every claim traceable to a real source. It is genuinely good at this: following a name from one census to the next, cross-referencing a map against a directory. Insist on sources. The story is better when it's true.

3  Chain-edit, don't generate

Here is the trick that makes the visuals hold together. Do not generate each era from scratch. Instead, take your real photo and ask an image model to edit it back into the previous era. Then edit that result into the era before. And so on, back through Victorian brick, farmland, forest, ice, sea.

Each frame inherits the structure of the last, so the house stays the same house and the land stays the same land, rather than becoming a different building every frame. We used Gemini 3 Pro Image via OpenRouter, then upscaled only the keepers.

4  Assemble the scroll

The page itself is simple: fixed full-screen image layers that cross-fade as you scroll, one caption per era. No framework required, though nothing stops you.

The gotchas, honestly

• Models drift. They love adding parked cars, putting a cheerful blue sky on everything, and quietly redrawing your building. Chain-editing fixes most of it; for the rest, prompt directly against the specific thing the model keeps sneaking in.
• Some frames are the wrong fit. Photoreal ancient organisms sent our model straight to CGI, and it lost the composition trying. Sometimes the answer is a different tool or a manual pass, not a cleverer prompt. Know when to stop arguing with it.
• Phones have limits. Dozens of large images will crash a mobile browser. Use a small pool of image slots and recycle them as you scroll, so only a handful exist in memory at once.
• Bad connections lie. Images arrive out of order and the wrong caption lands on the wrong picture. Never show an image until it has actually decoded, and hold the last good frame rather than flashing the wrong one.
• Cost is not the problem. Chain-editing runs to cents per frame; the whole project came in under about ten dollars. Generate cheap, upscale only what you keep, and it stays that way.

Go look

The record of an ordinary house is richer and stranger than you expect, and most of it is a search away.