Telling vs. Showing
Most people’s first instinct with an AI tool is to describe what they want in words: “write a product description that’s punchy and professional.” The model has to guess what “punchy” and “professional” mean to you, out of millions of plausible interpretations. Few-shot prompting skips the guesswork by including two or three concrete examples of the exact input-output pattern you want, directly inside the prompt, before asking the model to continue the pattern on new input. Instead of describing the target, you demonstrate it — and demonstration compresses ambiguity far more efficiently than adjectives do.
Why Examples Work Better Than Adjectives
Language models are fundamentally pattern-completion engines: shown a consistent structure, they’re very good at extending it. A single example (“one-shot”) helps, but it can’t distinguish what’s essential to the pattern from what’s incidental to that one case. Two or three varied examples (“few-shot”) let the model triangulate: it sees what stays the same across every example (the real pattern) versus what changes (the surface details), producing output that matches your intent far more reliably than adjectives like “concise” or “friendly,” which every person and every model interprets slightly differently.
What a Good Few-Shot Prompt Looks Like
A strong few-shot prompt has three parts: a short task instruction, two to four input-output example pairs formatted identically to each other, and then the new input with the output left blank for the model to fill in. The examples should vary in content but stay rigidly consistent in format — same length, same structure, same tone — because any inconsistency between your examples becomes a mixed signal the model has to resolve on its own. If you want short, two-sentence customer replies, every example should be exactly that length; if one example runs to five sentences, don’t be surprised when the model’s output length drifts.
When to Reach for It
Few-shot prompting earns its extra length whenever format and tone matter more than raw creativity: classification tasks, data extraction into a consistent structure, matching a specific brand voice, or converting messy notes into a standardized template. It’s less useful for genuinely open-ended creative work, where constraining the model to match prior examples can flatten originality you actually wanted. A practical habit: start with zero-shot (just the instruction) for a first draft, and only add examples once you notice the output drifting from what you actually need — that drift is usually the clearest sign an instruction alone wasn’t specific enough.