Set up a Claude agent to handle sushi orders

A developer used Claude to chat with customers, answer menu questions, and route confirmed Instagram orders for a seven-branch sushi business.

Done with Claude· Advanced

What they did
They linked Instagram DMs from Meta’s API to Claude Sonnet 4.6 and loaded branch-by-branch details for all seven restaurants. The agent had access to item lists, ingredients, calorie counts, allergen data, delivery areas, opening times, prep estimates, promos, and reply rules. In conversation, it can explain the menu, ask for the delivery address and item count, suggest extras, issue a bank payment link, and send accepted orders to the kitchen while recording them in the CRM and owner dashboard. Calls, voice notes, photos, and any direct request for a human still go to staff.
How it went
The developer says the system now deals with plain-text order chats almost end to end and gets the order right about 98% of the time. They also report that roughly 97% of the large prompt is served from cache, which reduces that part of Claude input cost by about 90%. The main concerns they called out were stale menu data and allergen errors.
Tips
Keep the menu, allergen list, and zone data current or the agent will drift. Make sure a human can take over quickly for non-text messages or any step that should not happen without approval.

Try it yourself with Claude

You are the chat-based ordering assistant for [restaurant name], a business with [number] locations. Use the data I provide on dishes, ingredients, calories, allergens, delivery coverage, opening hours, prep times, and promotions to answer questions, verify whether the customer’s address is inside a delivery zone, and gather the items and quantities they want. You may recommend add-ons and then draft the final order summary. Before you send a payment link, submit the order, or mark it complete, pause and ask the customer to confirm. If the customer sends a call, voice message, image, or asks to speak with a person, stop and hand the conversation to staff. Data: [paste menu, allergens, zones, hours, promos, and rules].

Fill in the [brackets]. Agents can make mistakes, so review anything before it is sent, paid or deleted.

Get 3 agent uses like this every morning

With the day's top AI news. Free, and one click to unsubscribe.