← all posts site ↗

inspire me pulls directions from my own vault

the visual bookmark app got an inspire me feature and it's the best thing it does now. the idea was simple: i have thousands of saved posts that are all things i liked, so when i start a new project i want the app to look at my own taste instead of generic inspiration. the implementation scores every embedded image against a style vocabulary locally — gradients, typography, posters, pixel art, dithering, ascii, 3d renders, the whole zoo — then picks a deliberately mixed set of references from different families and sends those plus their sampled palettes to a vision model. the output is a direction: palette colors with semantic roles, real font families, one cohesive materials language, tags, and the actual saved posts as reference chips.

the interaction settled into something good. a bolt icon beside the search box, and the same text field does double duty: enter searches, bolt inspires. a blank box means full mishmash across everything with a rotating lens so it's not the same output twice. seeding from a specific saved post works too, and the whole thing is available from chat through mcp.

two real lessons came out of it. first, quality control on the palette: the model once suggested red, cyan, and violet as near-equal accents, which is art, not an interface. the backend now runs contrast checks against wcag targets and a role sanity pass, and the ui shows a warning when a direction would break them. second, the embedding lesson — when i tried fusing image and text vectors into one embedding, the text side collapsed everything into the same region and search drowned in junk. the fix was keeping item vectors purely visual and folding text in only at query time. the model exists to suggest; the math exists to check the suggestion.