Zero, one, or five: how many examples

The three settings, and what each actually does

Zero-shot is a prompt with no examples. Just the instruction. This is what you use most of the time, and on current models it is far stronger than most prompting advice implies, because instruction tuning trained exactly this behaviour.

One-shot is a single demonstration. Its main job is not teaching a rule (one example cannot establish a boundary) but anchoring a format. Show one filled-in entry and you have communicated the layout, the level of detail, the punctuation, whether the field is a sentence or a fragment. Prose descriptions of layout are unreliable in a way that one filled example is not.

Few-shot is two to five. Now you can carry a boundary: the awkward case, the exception, the pair that shows where the line falls. This is where examples do the work prose cannot.

Beyond five, the returns fall away quickly for most everyday tasks, and the costs do not.

The costs, stated plainly

Space. Five worked examples of a document-analysis task can be two thousand tokens. That is two thousand tokens not spent on the document. In a long task this trade is real and it usually goes the wrong way.

Narrowing. Examples pull output towards themselves. If you want variety (names, taglines, approaches to a problem) examples are actively harmful, because you will get variations on what you showed. If you want conformity, this is the entire point.

Staleness. An example encodes a decision made on a day. Prices change, policies change, the category boundary shifts, and your prompt keeps demonstrating the old world confidently.

Cost per call. If a prompt runs a thousand times, five examples are five examples a thousand times over. This matters for anything automated, and not at all for a chat you send once.

A decision rule

Ask what kind of thing you are trying to transmit.

• A fact or a constraint? State it. “Reply in Portuguese”, “under 150 words”, “use only the pasted document”. No example needed, and an example would be a waste.
• A shape? One example. A filled-in row, a finished entry, a completed template.
• A boundary or a judgement? Two to five, including the hard case. Category edges, tone, what counts as too informal, which errors matter.
• Variety? None. Describe the constraint and leave the shape open.

Most disappointing example-heavy prompts are trying to use examples for the first category, where a sentence would have been clearer and cheaper.

The interesting middle case: format plus judgement

The commonest real task needs both. You want a specific output shape and a rule about what goes in it. The efficient version states the rule and shows the shape:

Summarise each complaint in one line: <product> — <what went wrong> — <what the customer wants>. If the customer does not say what they want, write unstated. Do not infer.

Example:

Blue kettle — arrived with a cracked lid — replacement, not refund

One example, one stated rule, one stated fallback. That will outperform five examples with no rule, and it takes a tenth of the space.

What changed with better models

Advice from 2023 recommended examples much more heavily, and it was right at the time. Weaker models needed demonstration for tasks that current models handle from an instruction.

Two things have not changed. Examples remain the best way to transmit a boundary you cannot articulate: the awkward case is still the awkward case. And they remain the best way to transmit your house style, because your house style exists nowhere in the training data and cannot be described accurately in adjectives.

So the modern pattern is fewer examples, chosen harder. Not a wall of demonstrations, but two that carry information nothing else could.

The check

Before adding an example, finish this sentence: this example teaches ___, which I could not say in a sentence. If you can complete it, keep the example. If you find yourself writing “it teaches it to be thorough”, delete the example and write a constraint instead.

That test also tells you when to stop adding. Once two of your examples would complete the sentence the same way, one of them is redundant, and redundancy in an example set is not harmless padding: it shifts the proportions and pulls output towards whatever those two have in common.