Zero-shot and few-shot: teaching by example

Sometimes describing what you want is harder than showing it. Try explaining the exact tone of your college's notice board messages in words, then compare that with simply pasting two old notices. The second works better, for people and for AI.

In this tutorial you will sort customer reviews for the Brew & Bean coffee shop into labels, first with instructions only, then with examples, and see where examples help. You need a free chat tool (ChatGPT, Gemini, Claude or Copilot). Python is optional and only used at the end.

"Shot" just means example. The model does not learn or change permanently from your examples; it uses them as patterns inside this one conversation. Start zero-shot. Add examples only when the answers are inconsistent or don't match what you had in mind.

Classify each customer review of a coffee shop into one label: taste, price, service or other. Reply with one line per review in the form: number - label.

1. The cold brew was perfect but I waited 25 minutes at the counter.
2. Rs 220 for a cold brew? I can get two filter coffees for that.
3. Banana bread was dry and tasted a day old.
4. Staff remembered my order, so sweet of them.
5. Do you deliver to Koramangala?
6. Great chai, but the price went up again.

What to expect: Most models get 3, 4 and 5 right. Reviews 1 and 6 mention two things, so you may get different labels on different runs or tools. That inconsistency is the signal that examples will help.

The problem is not that the model is wrong. Review 6 praises taste and complains about price; both labels are defensible. You have a rule in your head (the owner cares most about the complaint), and the model can't know it. You can write the rule down, but showing it is often clearer.

Classify each customer review of a coffee shop into one label: taste, price, service or other. If a review mentions more than one thing, label the complaint, not the praise.

Examples:
Review: Loved the cappuccino, but nobody answered when I called about my order.
Label: service

Review: Croissant was flaky and buttery, worth every rupee.
Label: taste

Review: Nice coffee, but ₹180 is a lot for a student.
Label: price

Review: Is there parking nearby?
Label: other

Now classify these. Reply with one line per review in the form: number - label.

1. The cold brew was perfect but I waited 25 minutes at the counter.
2. Rs 220 for a cold brew? I can get two filter coffees for that.
3. Banana bread was dry and tasted a day old.
4. Staff remembered my order, so sweet of them.
5. Do you deliver to Koramangala?
6. Great chai, but the price went up again.

What to expect: Expected: 1 - service, 2 - price, 3 - taste, 4 - service, 5 - other, 6 - price. The examples show the "label the complaint" rule in action, so the mixed reviews 1 and 6 now come out the same way every time.

Models copy everything about your examples, not only the part you meant. If all your examples are short, answers get short. If three of four examples are labelled price, the model leans towards price. If every example is about coffee, it may hesitate on a review about the sandwich.

Examples are just as useful for tone and format. Say Brew & Bean posts a short daily special on WhatsApp and Instagram. Instead of describing the style, paste two past posts.

Write a post in the same style as these examples. Match the length, the structure (item, one sensory line, price, pickup line) and the tone. Do not copy their wording.

Example 1:
Today's pick: Cold Brew. Slow-steeped overnight, smooth and chocolatey. ₹220. Order ahead, pick up in 5 minutes.

Example 2:
Today's pick: Masala Chai. Ginger, cardamom and a proper kick. ₹90. Order ahead, pick up in 5 minutes.

Now write one for: [item name, price, one fact about it].

What to expect: A good answer follows the four-part pattern exactly and stays about 20 words. Check that the price matches what you gave and that it didn't reuse "smooth and chocolatey".

If you classify reviews often, keep your examples in one place and generate the prompt. This script does not call any AI; it prints a prompt you can paste into a chat, or later send through an API (covered in *Your first LLM API call* in the AI Engineering track). Save it as build_prompt.py and run it with python build_prompt.py.

EXAMPLES = [
    ("The cold brew was perfect but I waited 25 minutes at the counter.", "service"),
    ("Rs 180 for a cappuccino is too much for a student.", "price"),
    ("Masala chai tasted exactly like home. Loved it.", "taste"),
    ("Is there parking near the shop?", "other"),
]

new_review = input("Paste a review: ").strip()

lines = [
    "Classify a Brew & Bean customer review into exactly one label:",
    "taste, price, service or other. Reply with the label only.",
    "",
]
for text, label in EXAMPLES:
    lines.append(f"Review: {text}")
    lines.append(f"Label: {label}")
    lines.append("")
lines.append(f"Review: {new_review}")
lines.append("Label:")

print("\n".join(lines))

Ending the prompt with "Label:" invites the model to complete just the label, which keeps answers short and easy to read in code.

For AI and data roles, a common question is: "How would you get an LLM to classify support tickets into our categories?" A good answer: start zero-shot with clear label definitions, measure on a small labelled set, add a few varied examples for the categories it confuses, and only consider fine-tuning if prompting can't reach the accuracy you need (compared in *Prompting vs RAG vs fine-tuning* in the AI Engineering track).