Michael Kilkelly from ArchSmarter joined me for Episode 139 to show how AI has changed what one person can build for Revit. The short version... a custom design tool that used to take a month of hand coding now takes an afternoon. If you have ever wanted a tool that fits your exact workflow but figured it was too much work, this one is for you.
What We Covered
Michael's whole approach runs on one idea he calls the Design Intent Pattern. Architects control design intent. Contractors handle means and methods. He treats AI the same way. He owns what the tool should do, then lets the AI handle how it gets coded. That framing alone is worth the watch, because it reframes AI from "thing that writes code" into "thing that handles the part you do not need to care about."
The image-to-family demo was the one that got people talking in the chat. Michael dragged a photo of a bench into Claude, and it generated a structured data file describing the geometry. A small Revit tool then read that file and built a native family. Not perfect, not parametric, but close enough to be useful. The interesting wrinkle... AI rebuilt a 3D object from a single photo, but it still struggles to read a flat 2D floor plan. As Michael put it, certain spatial relationships just trip it up.
We also went through his mini design tools. A twisting tower generator he built at his kitchen table on a Sunday. StairLab, a single HTML page that lets you lay out stairs visually and checks them against IBC 2021 as you go. And Charrette, a rendering tool that grabs a Revit screenshot, sends it to the Gemini API with a prompt, and sends back a finished image. Each one is small, cheap to build, and basically throwaway. That is the point.
The honest part lands near the end. The value now is not syntax or where to put a semicolon. The value is domain knowledge. Knowing what the problem is and how it should be solved. AI handles the rest. We also got practical about deploying this stuff at scale, code review, and managing the cost of running multiple models.
