Use eight prompting terms to improve one request. The job is to turn a customer’s delivery message into a support triage row. Keep the message, routing rules and scoring fixed so you can see what an added example changes.

Build a request you can score

  1. A prompt says what to do. Ask for a queue and a status field instead of “handle this ticket.”
  2. Context gives the model the facts it can use. Supply the customer’s message and the approved routing rules in separate blocks.
  3. A prompt template holds the stable directions and a slot such as [customer message]. Replace the slot for each ticket and keep the field names consistent.
  4. Prompt engineering is designing and evaluating a request. Check the first draft against known cases, revise one part and score the same cases again.

Compare two versions of the same request

Zero-shot example

Zero-shot prompting gives the task and rules without a demonstration. Send this template with the message filled in.

Route the message to Shipping for late or missing packages, Billing for payment issues, or Other for everything else. Use the delivery date only when the customer explicitly gives it. Otherwise write “Not stated.” Return one line in the format Queue | Delivery date.

Customer message: “The package for my order hasn’t arrived. Can you look into it?”

Few-shot example

Few-shot prompting shows the model how to carry out the task. Keep the same brief and target message. Insert these demonstrations between the rules and the target message.

Message: “The order was due on May 12 and hasn’t arrived.”
Row: Shipping | May 12

Message: “My parcel is missing. I placed the order on May 9.”
Row: Shipping | Not stated

Check what the model returns

The output is the returned result. The expected example is Shipping | Not stated. Score both versions against the same three checks.

  • One line with two fields in the requested order.
  • Shipping because the package hasn’t arrived.
  • Not stated because the message gives no promised delivery date.

Where application instructions fit

A system prompt supplies application-level instructions, such as a support app’s approved queue names. The term describes the purpose of the instructions. A system message is a role name some APIs accept. In OpenAI’s Responses API, developer is the documented role for rules from the app builder, while user carries the end user’s request. The developer instructions take precedence over the user’s instructions. A chat interface may expose only the user message field. Instructions and demonstrations guide the current response. Learned model parameters are values set through training. OpenAI’s text generation guide explains the roles, and its Responses API reference lists the accepted message types.

Use the course exercise on prompting with examples to try the technique on a new input. Include a payment issue and a message with an order date but no delivery date when checking a routing template.