Zero-Shot Prompting
The simplest way to use a model is to describe the task and let it go, with no examples at all. For Spoke & Chain's triage that can work surprisingly well — or produce labels like "Bicycle Maintenance Inquiry (Positive Sentiment)". The difference lies entirely in how precise the instruction is.
The rule
No examples means the instruction must carry everything.
A zero-shot prompt gives a task and no worked examples. The model relies on what it learned in training about tasks like this. That is often enough for common jobs — summarise, translate, classify into obvious groups — because modern models have been trained heavily to follow instructions.
With no examples to copy, though, every detail you leave unsaid is a guess. A strong zero-shot prompt pins down:
- The exact task — "classify", not "tell me about".
- The allowed answers — a closed list, when there is one.
- Definitions for borderline cases — what counts as a refund?
- The input boundary — delimiters around the data.
- The output format — exactly how the answer should look.
Classify the customer email between the <email> tags into exactly one
category: repair, order, refund, other.
- refund: the customer wants money back, even if other topics appear.
- order: buying, tracking or changing a purchase.
- repair: fixing or servicing a bike.
- other: anything else.
Reply with the category only, in lowercase.
<email>
The bell I bought is cracked. Can I just get my money back?
</email>Expected output: refund.
The shortcut
Read your prompt as if you were a new temp worker on day one with no one to ask. Every question you would need to ask is a gap to fill.
"What kinds are there?"
Name the four allowed categories. The model can no longer invent its own.
Zero-shot is the right place to start: it is short, cheap and easy to change. If the test set still shows the model misunderstanding the pattern or format, that is the signal to add examples, which is the next lesson.
Drills
Answer fast — trust the rule.
What makes a prompt zero-shot?
Which zero-shot instruction is strongest for the triage task?
The prompt labels "fix my wheel and refund the bell" inconsistently. What is the best zero-shot fix?
When is it time to move from zero-shot to few-shot?
Remember
- No examples means the instruction must carry everything.
- Shortcut: read it as a day-one temp worker and fill every question.
- Trap: vague tasks with open-ended answers and no output format.