Projects Cooking Miner
The AI that rates my recipes never gets the publisher field.
A recipe miner that asks how excited I’d be to cook a dish, and a thin day that ships a short list instead of filler.
- Title
- Slow Braise with Prunes and Bay
- Ingredients
- Short ribs, onions, prunes, bay leaves, red wine, black pepper
- Instructions
- Brown the ribs. Soften the onions in the fat. Add everything else, cover, and braise until a spoon goes through.
Publisher- Removed before the model sees the recipe:
Example Kitchen
Every night this thing reads recipe sites, and newsletters get read as they land. A model scores what it finds, and I get a daily list aimed at five dishes. The publisher field is left out of what the model receives, and a test checks that it’s missing. The prompt also says not to reward or punish a publisher, but I don’t rely on a prompt for that. Scraped page text can still mention a name, so this removes the field, not every trace.
The first version had a side door. A recipe too old for the normal list could qualify for the wildcard slot at 55 if a publisher had recently featured it in a newsletter. The rewrite took that out. A feature now earns one sentence beside the recipe and, if its score is older than the feature, a fresh score and a new window on the list. Nothing in the ordering looks at it.
That version also treated recipes like news. It dropped anything over two weeks old and ranked every fresh recipe above every older one, whatever the score. A braise doesn’t go stale. The nightly crawl still only picks up recent posts, but once a recipe is in the pile, its publication date stops counting, so an old 72 beats a fresh 52. Agents write this code, so a test is named for exactly that.
Nearly every recipe online is competent, so the score asks how excited I’d be to cook it this week. The prompt tells the model to spread its scores, because a cutoff can’t cut a pile. The model never sets the flag that says recommended. Code does, then checks again at selection. The one way onto the list without that flag is a single wildcard scoring 60 to 69, labeled on the card.
A demo always shows five good cards. A real Tuesday might only have two. Recently shown recipes can come back to reach three, and the card admits it. Past that, it would rather give me two recipes and raise an alarm than pad the list to five.