the old way of referencing my visual inspiration was stupid. i'd tell the assistant to look at my shooting style and it would run a search, get a handful of results, then fetch the image from each result one by one just to answer. fine at five results, awful when the whole point is "see the entire wall at once."
so i added a tool to the taste engine mcp called browse_vault. one call takes a query and runs the same search the site uses, then renders every match as a numbered grid — up to sixty tiles, pinterest wall style, video badge on the videos, hue and kind labels — and returns the whole sheet as one image with a legend and a table mapping each number to its item id. to study a specific tile you follow up with get_vault_images for just that one. one round trip instead of one per result.
tested it properly before calling it done. a unit render with thirty six tiles, then a real call over streamable http for "scenic mountain drive" came back as seventeen tiles in a single 78kb image. committed and pushed, with the rendered sheets cached to disk and only the last twenty kept.
now "look at my shooting style inspiration" means the agent actually sees fifty shots at once and answers with the whole wall in context instead of sampling. the job went from a pile of round trips to one.