Turn a single floor-plan image into a Mapbox indoor map

Akira Murao built a personal tool, indoor-generator, as a repo of Claude Code skills plus a fixed Python pipeline. Claude Code does the research, writes a JSON config per building and talks to the user; the pipeline turns that config into Mapbox indoor GeoJSON. For the 5th floor of PMO Shinbashi II in Tokyo Murao handed over one rental floor-plan image and a one-line request. Claude found the building in OpenStreetMap, looked up floor count and floor area on the web, drew the footprint from the plan because the OSM outline was out of date, scaled and rotated the image onto the building, traced the rooms and pictograms, then built the GeoJSON and started a local preview. Murao also ran it with no image at all on an outlet mall, where Claude found the official floor guide images itself.

The office floor reached the preview about five minutes after the request: 16 rooms, 15 of them traced from the image and covering about 49% of the area, plus 6 points for lifts, stairs and toilets. Room names came from the text on the plan. Toilet stalls and curved walls were simplified, and which way was up had to be judged from surrounding roads. The mall took many rounds of chat fixes, about ten of them store by store, before its 125 store outlines over two floors looked right. A first attempt at warping images with control points did not work and was dropped.

Accuracy is a few metres, fine for event or office maps but not for positioning or routing. The tool keeps track of which rooms were traced, estimated or derived so you can see what to trust. Floor plans from official sites carry copyright, so check the rights before publishing. The post sits on Mapbox's dev.to page but the author calls it a personal project, and the figures were measured on an earlier version of the tool.