The Instruction That Never Quite Lands
You’ve probably tried this: you write a detailed paragraph telling an AI tool exactly what format you want — tone, length, structure, style — and the output still comes back slightly off. Not wrong, just not quite what you pictured. The fix often isn’t a longer, more precise paragraph. It’s abandoning the paragraph altogether and showing the model two or three examples of exactly what “right” looks like. This technique is called few-shot prompting, and it works because of how these models actually process information.
Zero-Shot vs. Few-Shot
A “zero-shot” prompt is just an instruction with no examples: “Summarize this email in one sentence.” It often works fine for simple tasks. A “few-shot” prompt adds a small number of worked examples before the real request, showing the input-to-output pattern directly: here’s an email, here’s the one-sentence summary; here’s another email, here’s its summary; now summarize this one. The model isn’t just following a rule anymore — it’s pattern-matching against demonstrations, which tends to produce more consistent formatting, tone, and length than a description of those things ever does.
Why Examples Beat Descriptions
Language is lossy. When you write “keep it casual but professional,” you and the model may have different internal pictures of what that means. An example removes the ambiguity entirely — it’s not a description of casual-but-professional, it is casual-but-professional, in a form the model can directly imitate. This is especially powerful for anything with a specific structure: a particular JSON shape, a consistent report format, a tone you’ve used in past emails, or category labels that don’t map to common definitions. Two or three good examples often out-perform a full page of written rules for these structured tasks.
How Many Examples You Actually Need
More isn’t always better. One example (one-shot) can be enough to lock in a format. Two or three examples (few-shot) is the sweet spot for most everyday tasks, since it lets the model see what varies between examples versus what stays constant. Beyond four or five examples you usually hit diminishing returns, and you start burning context space that could go toward the actual content. If your examples disagree with each other on formatting, the model will pick up on the inconsistency and may blend or alternate between them, so consistency across your examples matters more than quantity.
Putting It Into Practice
Next time an AI tool keeps missing the mark on format, resist the urge to add another sentence of explanation. Instead, paste in two real examples of input paired with the exact output you want, then add your actual request below them. For recurring tasks — weekly reports, email replies, data extraction — save that example pair somewhere so you can paste it in every time rather than re-explaining the format from scratch. The few minutes it takes to assemble good examples usually pays for itself the very first time you reuse them.
