Zero-Shot: Just Ask

The simplest pattern in the toolkit. Zero-shot means you describe the task in words, give the model the input, and trust it to do the work without showing any examples. Most prompts you wrote in Modules 1 and 2 were already zero-shot.

Module 3 walks through six core patterns. Here is the order:

When zero-shot works

Modern models like Claude Sonnet 4.5, GPT, and Gemini are strong enough that zero-shot is the right starting point for most everyday tasks. Try it first. If the output is good, you don't need anything more elaborate.

Zero-shot tends to work well when:

• The task is common (writing an email, summarising a paragraph, classifying obvious categories).
• The categories or shape of the output are already well-known to the model.
• The instruction itself is clear and bounded.

A clean example

Watch this prompt — a simple classification task. We want to bucket a customer email into one of four categories.

Classify this customer email as one of: billing, technical, sales, other.
Reply with only the category.

Email: "Hi, my invoice for March shows a $40 charge I do not recognise. Can you check?"

Captured output:

billing

Captured from Claude Sonnet 4.5 (claude-sonnet-4-5) on 2026-04-27. Re-runs may differ slightly.

That's the entire response — one word, exactly as asked. No preamble, no "Sure, here's the classification:", no markdown. The output spec ("Reply with only the category") locked the format, and the categories were obvious enough that no examples were needed.

When zero-shot starts to fail

You'll know you've hit zero-shot's limits when:

• The model picks a different category every time, even on similar inputs.
• The model hedges and explains instead of giving a clean answer.
• The output format drifts (sometimes lowercase, sometimes title case; sometimes with punctuation, sometimes without).
• The categories are domain-specific and the model isn't sure what each one means.

That's the moment to upgrade to few-shot (next lesson) — you stop describing the task and start showing it.

A small habit

When you sit down to a new task, do a quick zero-shot pass first. It takes 30 seconds and answers a useful question: is this task in the model's comfort zone, or am I going to need to teach it?

If zero-shot looks roughly right, polish the prompt with a sharper output spec and ship it. If zero-shot is wobbly, that's your signal to add examples.

What zero-shot is not

Zero-shot does not mean "minimal prompt". A zero-shot prompt can still be 200 words of clear specification. The "zero" refers to examples, not to detail. A well-specified zero-shot prompt with role, task, context, and output spec can outperform a shoddy few-shot prompt with bad examples.

Next: few-shot — when to teach the model with 2 to 5 worked examples. :::