AI · 2026
Stockholm Palace, built by Claude in SketchUp
I asked Claude to build Stockholm Palace in SketchUp without opening the program. The first version was based on Wikipedia and looked like an office complex. Two drawings from 1897 showed the courtyard was turned a quarter turn, and that the hard part was not building but deciding which source was right.

I have used SketchUp for a long time. When we renovated, I measured the whole flat and built it in the program, room by room, and dropped in the manufacturers’ CAD files for furniture, lamps and beds to see whether everything really fit. So I know roughly how long it takes to build something in SketchUp. The question was what happens if I do not build it myself.
SketchUp now has a connection that lets Claude work directly in a cloud model, without the program being open on my computer. I asked for a model of Stockholm Palace. No drawing, no file, just the name of the building.
The first version
Claude found the dimensions on English Wikipedia and built from them. The result was a rectangle around a courtyard, with four wings and two curved wings, at the right size and in the right place. It looked like an office complex.
So we kept going, the way you do in SketchUp: cornices and plinths dividing the façade into storeys, windows with stone surrounds instead of holes in the wall, terrain with Lejonbacken to the north and Logården to the east. It got better with every pass. But the proportions felt wrong, and I could not say why.
Better sources
So I asked for more sources, and that changed everything. Wikimedia Commons holds two drawings of the palace that are old enough to be free to use: a plan with measurements from 1897 and a floor plan of the ground floor.
Claude did not look at the drawings. It measured them. The scale bar was found as the longest run of black pixels in the image. The building volumes were separated from the paper with image analysis. The curved wings were worked out by fitting a circle to thousands of edge points and reading off where the edges actually were. Out of that came dimensions to two decimals, which could be compared with what was already in the model.
The difference was not small. Wikipedia had swapped the two sides of the courtyard: 89 by 77 metres, where the 1897 drawing says 77.64 by 88.92. The whole main building was turned a quarter turn. Claude had guessed the curved wings at radii between 10 and 26 metres, hung on the gables of the wings. The real ones are 31 to 42 metres, drawn around a point 82 metres west of the centre of the courtyard. The total length of the palace went from 244 metres to 229.4. The real figure is around 230.
This was where I understood what the project was really about. The hard part was not getting the tool to build geometry. The hard part was deciding which source was right when they contradicted each other. An encyclopaedia written by many, or a drawing made by someone who stood on site with a tape measure in 1897. A tool cannot answer that question. Someone has to ask it.
What went wrong
The session against the cloud model crashed several times, mid-build, and everything had to be redone from the last saved version. The connection’s own preview is 512 pixels wide, which is enough to see that something is there but not to judge whether it looks right. So Claude wrote a small renderer of its own to deliver images at higher resolution. The first version drew the surfaces in the wrong order, so walls far away ended up in front of walls close by. And the first time the camera was set up, half the palace was out of frame. None of that shows in the finished result, but it took a large share of the time.
The result
A SketchUp file with around 10,500 surfaces, organised in a named hierarchy so it can be worked on further: ranges, wings, curved wings, terrain. The 777 windows are components, so changing one changes them all. Plus four renderings, one from each point of the compass, at 2,800 by 1,750 pixels.
The model is far from correct. It is still a massing model: no dormers, no balustrade, no lion on Lejonbacken, and anyone who knows the palace will find errors everywhere. But the dimensions match a drawing from 1897, and it is closer to the truth than any living person would get in the 30 minutes it took to produce, however good at SketchUp they are.
That is exactly what working with AI is about. Making sure what goes in is right, and judging what comes out for what it is. It is not a magic machine that understands everything you do, like a best friend. It needs clear, structured information. And you need to know what it is about.




