The poster built a Claude agent that takes orders for a seven-location sushi chain, using menu, allergen, delivery zone, hours and promo data.
The poster built a Claude agent that takes orders for a seven-location sushi chain, using menu, allergen, delivery zone, hours and promo data.
The builder wired Instagram DMs from Meta’s API into Claude Sonnet 4.6, giving it a knowledge base for all seven locations: menu items, ingredients, calories, allergens, delivery zones, hours, prep times, promos, and conversation rules. The agent chats with customers, answers menu questions, collects address and quantity, offers add-ons, sends a bank payment link, then pushes confirmed orders to the kitchen and logs them in the CRM and an owner dashboard; humans still handle calls, voice notes, photos, or explicit requests for a person.
The poster says the agent handles plain-text order chats almost entirely on its own, gets orders right about 98% of the time, and rarely needs human takeover beyond non-text cases. They also say roughly 97% of the large prompt is served from cache, cutting that portion of Claude input cost by about 90%, but allergen mistakes and stale menu data were raised as the key risk.
The main practical catch is keeping menu and allergen data current; the post says the economics only worked because Claude prompt caching avoided reprocessing the full seven-location context on most messages.
Act as an order-taking agent for [restaurant name], which has [number] locations. Use the menu, allergen info, delivery zones, hours and promotions I paste below to take customer orders, ask which location they want, warn about allergens, and confirm the delivery address is in a zone. Finish each order by reading back the items and total, and ask the customer to confirm before it's marked as placed. Here is my data: [menu, allergens, zones, hours, promos].
Score every demo request against your ICP, update HubSpot and ping the right rep in Slack within a minute of the form submit.
Every new Zendesk ticket gets a category, priority and sentiment tag plus a draft reply in an internal note before an agent opens it.
Copilot and Claude review each PR for bugs, missing tests and risky changes before a human looks, so reviewers can focus on design.