Select exactly {{num_recs}} products from the provided catalog for this shopper. Goal: - Maximize shopper fit and recommendation usefulness. - Build a coherent set that matches current intent and can improve conversion. - Use strong catalog evidence only and avoid weak guesses. Ranking policy: - Start from the viewed item when present and use category breadcrumbs, function, style, materials, use case, season, and price band. - Use cart and order history to infer the shopper mission, avoid repeats, and prefer complements, compatible add ons, or close substitutes. - Use persona only as a tie breaker. - Prefer exact category and attribute matches over broad semantic similarity. - Treat breadcrumb separators and detailed product names as important signals. - In a large catalog, first narrow to the most relevant category neighborhood, then choose the strongest unique items. - Avoid the same base product in another color, size, scent, flavor, pack, or bundle. - Keep gender consistent if the viewed item is gendered. - Never confuse pet products with baby products. - If the viewed item is NA or missing, infer intent from cart and order history and choose the safest high precision matches. - If evidence is weak, choose conservative relevant items rather than creative ones. Data: Persona: {{persona}} Viewed item: {{sku}}{{sku_info}} Cart: {{cart_json}} Order history: {{order_json}} Catalog: {{product_catalog}} Output rules: - Return only a JSON array. - Return exactly {{num_recs}} objects sorted from strongest to weakest recommendation. - Each object must contain exactly these keys: "sku", "name", "price", "reason". - Copy sku, name, and price from the catalog exactly. - Every returned item must exist in the catalog. - Do not return the viewed item. - Do not return items already in cart or order history. - Do not return duplicate SKUs or near duplicate variants. - Keep reasons customer facing, concrete, and benefit focused. - Each reason must be one short sentence using plain words and no punctuation. - Do not mention internal strategy, business goals, scoring rules, or hidden reasoning. - Do not output markdown, code fences, commentary, or any text outside the JSON array. Final check before responding: - exactly {{num_recs}} items - valid JSON - unique items - all items copied from catalog - no viewed item - no cart items - no prior order items

bitrecs_haystack_daily