#Zero-Shot, One-Shot & Few-Shot Prompting¶

The single highest-leverage technique in your toolkit: showing the model what "good" looks like.

Zero-Shot¶

No examples — just an instruction. Relies entirely on the model's pre-trained ability.

1Classify the sentiment as positive, negative, or neutral.
2
3Review: "Battery life is disappointing."
4Sentiment:

Great for simple, common tasks. Fragile for anything with a specific format or edge cases.

One-Shot¶

Provide one example to anchor format and behaviour.

1Review: "Love it, works perfectly!"
2Sentiment: positive
3
4Review: "Battery life is disappointing."
5Sentiment:

Few-Shot¶

Provide several examples — especially ones covering edge cases.

1Review: "Love it, works perfectly!"
2Sentiment: positive
3
4Review: "It's okay, nothing special."
5Sentiment: neutral
6
7Review: "Stopped working after a week. Avoid."
8Sentiment: negative
9
10Review: "Battery life is disappointing but the screen is gorgeous."
11Sentiment:

That last mixed-signal example teaches the model how you want ambiguity handled — something instructions alone struggle to convey.

Why Few-Shot Works¶

The model continues patterns (Module 1). A consistent set of examples is a pattern. The model infers the implicit rule and applies it to the new input.

Best Practices¶

Diminishing returns: 2–8 examples usually captures most of the gain. 50 examples rarely beats 8 — and burns tokens.

When Few-Shot Isn't Enough¶

If even good examples fail, the task likely needs reasoning, not just pattern-matching. That's Module 3 (chain-of-thought).