What is scratch-food-rubrics?

scratch-food-rubrics is a set of skills meant to be used by AI coding agents like Claude Code or Codex, designed to help find restaurants and bakeries that actually make their food from raw ingredients on site, rather than reheating or finishing off pre-made components. It ships as two separate rubrics, one for restaurants and one for bakeries, each triggered by typing a slash command followed by a location. Each rubric scores candidate places on three measures: how much of the food is genuinely made from scratch, how interesting the place is to someone who eats or shops there often, based on menu variety and how rare that kind of food is locally, and a basic quality rating used only to filter out poor places rather than to rank them. The bakery version is tuned to catch specific tells of shortcuts, such as frozen dough that gets partially baked elsewhere and finished in store, which the README calls par-bake. The process the agent follows is fairly involved. It starts by creating a dedicated folder for that specific search so nothing overlaps between runs, then searches a general map data source for an unbiased list of nearby places, followed by more targeted, language aware searches to catch well known, new, renamed, or oddly labeled venues that map data might miss. Candidates are merged and deduplicated, then separate worker agents go out and gather quotes and evidence from menus, websites, reviews, and press coverage without being allowed to score anything themselves. The main agent then checks that evidence, asks for more detail if something is missing, applies the scoring rubric, and produces a final result organized by categories like standout finds and rare discoveries. A notable idea in the README is watching for what it calls involuntary tells, meaning physical details a business reveals without meaning to, such as a bakery selling out unusually early, which can contradict marketing language claiming everything is baked fresh. To use this project, you install or link the entire folder structure, not just a single file, since the skills share reference files. It is released under the MIT license.