I've been saving artefacts I respond to on Google Arts & Culture for 15 years: paintings, maps, furniture, architectural drawings. I wanted to find out whether that collection could tell me something useful about my taste, and whether I could turn it into guidance for things I make.
This is that experiment. The aim is a set of design.md playbooks for
interfaces, interiors and a wider design language, grounded in the works
I keep returning to.
Work in progress: 801 collected works, 800 with generated visual analysis, and ten provisional groups. The playbooks are still to come.
The first attempt was a sketch typed into AI Studio on my phone. It produced CuratorMD, an app with artwork analysis, curatorial themes and a design-token sandbox. My next prompt restyled it. A later attempt to extend it hit the model quota.
Looking back at the code, I found a more basic problem: the analysis had been given the artwork's title and metadata, but not the image.
I picked the idea up again as a pipeline. This time the model received the images, with a prompt asking it to consider composition, utility, symbolism, space and colour. Each stage became a chance to try a different approach, inspect the result and work out what to keep.
One generated observation about Fernando Botero's Arcángel suggests using a dominant foreground form against a repeating background to create depth. That gives me something specific to try in an interface, and a question: does the arrangement help someone find what matters?
Read the model's analysis · See the work at its source
That translation is the part I still need to test. An observation about a painting is a starting point, not a proven rule for designing software.
The catalogue contains the generated critiques and design suggestions. The cluster map is an exploratory view of ten groups, whose methodology remains under review. Some apparent connections may reflect artist names and the model's writing as much as visual relationships.
I selected the collection, set the brief and chose between approaches. AI tools generated much of the code and analysis. The critiques are the model's words; the notes also contain scripted text and space for my own annotations. Those annotations and a worked design example are next. The original app and recovered pipeline scripts are experimental, with local dependencies, rather than a ready-to-run service.
I'm opening up the working record while the project is unfinished. I want to make the work visible, share what I'm learning and give myself a reason to keep going. The false starts belong here too.
Follow the journal and next steps, browse the collection spreadsheet, or read the longer project history.
Artwork images and museum-supplied descriptions are not included. Catalogue notes link to their sources. Remaining museum-supplied metadata, such as titles and holding institutions, retains its own rights and is excluded from this project's content licence. Code is MIT; original writing and generated analysis are covered by the content licence.