isoscapeisometric maps of real places

Open this in the explorer

Questions and answers

Drafted with Claude; see the artist statement.

Claude and the code

Does Claude handle the projection, or the map geometry too?

Claude wrote all of the code, under my direction, and the code does both. The geometry doesn't come from Claude, though. Streets, water and building outlines are OpenStreetMap. The ground is USGS lidar. In Richmond the building dates come from the city's year-built records, and the lost buildings are traced from Sanborn fire-insurance maps at the Library of Congress. On Gwynn's Island some buildings come from Microsoft's machine-learned footprints.

The projection is a small part of the code. It's the camera step, view.ts, about 90 lines out of some 14,000. The rest reads what each OpenStreetMap tag means, guesses heights nobody mapped, finds the river's surface and rapids in the lidar, raises the bridges, works out what hides what, and paints the ground in strips from the back of the frame forward.

The isoscape pipeline. fetch downloads OpenStreetMap and elevation into a local cache. build then runs eight steps, each named with its file: osm.ts, classify.ts, history.ts, view.ts, terrain.ts, scene.ts, order.ts and render.ts. The camera step, view.ts, is outlined in red as the projection.

Builds are deterministic. The same data and the same commit give the same SVG, byte for byte. build doesn't use the network, and no model is asked anything while a map is drawn.

Is any of this AI-generated imagery?

No. No image model draws anything. You can rebuild any map from its commit and its data and get the same file.

A model, or Claude's judgment, does show up in the data in four places.

There's more in Measured, inferred, guessed and Time.

What do I do, and what does Claude do?

I decide what to draw and what it should look like. Claude writes the code, runs it, and shows me the result. I keep it or send it back. The code computes the drawing.

Who does what, in three stacked boxes. Chris Busse decides what to draw, writes the rules down in CLAUDE.md, and keeps each result or sends it back. Claude, under his direction, writes and changes the code, runs the builds and checks, and shows him what changed. isoscape, the code, turns the data into the SVG the same way every time, with no image model and no network. Arrows go down (a request, a build) and back up (pictures, an SVG with counts).

A lot of my part is writing down what a good map looks like. Those rules are in CLAUDE.md, so they carry from one session to the next. I wrote more about working this way in the artist statement.

What happens when something is wrong?

It gets fixed in the code and the map is drawn again. A print that's already for sale is never changed in place. It's redrawn as the next version, and the Changes page says what changed and when. Version 4 of Richmond, 1925 put back a missing span of the Mayo Bridge and took out traced buildings that were standing in the river.

Measured, inferred, guessed

What is measured, what is inferred, and what is guessed?

Where the geometry comes from. Seven sources, each with what it gives and a mark: solid for measured, a ring for inferred, a dashed ring for guessed. OpenStreetMap: streets, rail, water, parks and outlines, measured; tagged heights, measured; other heights, guessed. USGS 3DEP lidar: the ground, measured; river surface, rocks, rapids and shoreline, inferred. Microsoft footprints and NLCD land cover, on Gwynn's Island: inferred. Richmond's year-built records: measured. Sanborn maps: lost buildings, inferred; their heights, guessed. Bridge deck heights: guessed.

Each map's "How this view was made" tab gives its own count.

Why do some buildings look the wrong height?

Their heights are guesses. A building with no height or building:levels tag gets 8 m for a house, 16 m for apartments and 13 m for an office (Buildings). Real heights from lidar are first on the roadmap.

The historic years

How accurate are the historic years?

They're rough. Buildings are missing, some sit a few meters off, and every lost building's height is a guess from its material.

A dated Richmond map has four kinds of thing in it:

Downtown in 1900 had far more buildings than the 1900 map shows. Assessors round, so a lot of buildings are dated 1900 or 1910. The ground is today's lidar in every year, so the I-95 cut is there in 1737. A traced building that couldn't have stood where it lands is left out whole. In 1925 that's 27 in the water, 61 on viaducts or track, and 62 under ground that's higher today.

Time has the whole method. Fixes go on the Changes page.

How a map gets made

How does a map get made, step by step?

  1. Write a region file, regions/<name>.json, with a center, a width and a depth in meters.
  2. fetch downloads OpenStreetMap and the elevation into data/<name>/. Nothing else touches the network.
  3. build draws the SVG from that cache.
  4. Look at it in the explorer, out/index.html, next to the last build.
  5. For a print, add the title, credit line and title block, write the release record, and run the rebuild check.
node bin/isoscape.js fetch richmond-downtown
node bin/isoscape.js build richmond-downtown-topo --png

Why isometric, and why dots and lines?

Isometric lets you take in a whole neighborhood at once. The dots and lines come from my other project, dotscene. Every edge is an inked line, and every corner gets a dot. I got the idea of an isometric city drawn from open map data from Justin Chavez's Isometric SF.

Making your own

Can I make my own map? What do I need?

Yes. You need Node 22.18 or newer, and rsvg-convert if you want PNGs. There are no other packages to install. Write a region file for your place, then run fetch and build. The README has the commands and the settings.

Terrain works anywhere. It's USGS lidar in the US, national models in France, Switzerland and Japan, and Copernicus everywhere else. The code is MIT-licensed, and a map you make is yours. Credit OpenStreetMap wherever you show it (About).

How do I use it with Claude or another coding agent?

Open a chat with Claude and say "Use the GitHub repo busse/isoscape to make an isometric map of your place." It clones the repo, writes the region file, fetches the data and builds the map. That takes a few minutes.

Using it with an agent, in three stacked boxes. You name the place and look at the picture. Claude, in the app or a terminal, reads CLAUDE.md and AGENTS.md. The repo: the region file, fetch, build with --png, the explorer at out/index.html, npm test, and the visual check.

The rules for working in the repo are in CLAUDE.md. They cover the conventions, where to change what, the known limits, and the loop. AGENTS.md has notes for agents helping someone use isoscape. Codex, Cursor and other agents that read AGENTS.md get those too, and GEMINI.md points Gemini to it.

The loop says never to judge a change from the numbers alone. Build it, look at the picture, and compare it with the last build (press C in the explorer) before you keep it.

Say what you want to see and ask for a picture back. "The bridge in the middle ends in the air" is enough to start with.

How is the codebase kept in order when an AI writes most of it?

With rules written into the repo, and checks that enforce them.

What data does it use, and under what licenses?

The code is MIT. My published maps are © Chris Busse under CC BY-NC 4.0. You can share and adapt them with credit, but you can't sell them. Every credit is in About.

Prints

What am I buying when I buy a print?

A print of one of my maps, printed when you order it. The file it's printed from has a release record, so I can make it again exactly. Prints are in the holiday shop at chrisbusse.com through December 31, 2026, and buying one supports the project.

From drawing to print, in seven steps: build the scenes, look at them at print size, compose the print, add the title block with the plate and version, write the release record, run the rebuild check, and print it when someone buys one. An arrow runs from the last step back to the first. If a fault turns up after release, the design is redrawn as the next version, the plate stays the same, and the Changes page says what changed.

What does the title block on a print mean?

It's the small block in the bottom-right corner. It has the isoscape cube, "drawn by Chris Busse", the collection, set and piece, the plate number, the version, the year it was drawn, the point the view is centered on, and the direction it faces (Plates and the title block).

The plate never changes. RVA·01 is Richmond, 1925. The version goes up each time I redraw it.

What happens if a map I bought is corrected later?

Prints ordered after the fix use the new version. Yours has its version printed on it, so the Changes page will tell you what's different.

Open this in the explorer