Examples Beat Instructions — The "Show, Don't Tell" Rule in Prompting

You've written a careful prompt with detailed instructions and the output still comes back wrong — wrong format, wrong tone, wrong length. The fix is usually not more instructions. It's an example.

The jargon

Few-shot prompting — giving a model one or more input/output examples before your real request. "Few-shot" means a handful of examples; "zero-shot" means none.

One-shot — a single example. Often enough to shift format or style reliably.

The lesson

Words like "concise", "professional", or "structured" mean something different to Claude than they do to you. An instruction says what you want. An example shows it. Claude will pattern-match to the example rather than interpret your description of it.

This matters most when the output has a specific shape — a report section, a client email, a JSON object, a bullet format — and "close enough" isn't good enough.

How it works

Here's a zero-shot prompt that usually disappoints:

Summarise the following meeting notes in a professional, concise format.

[notes]

And here's the same request with one example:

Summarise meeting notes in this format:

INPUT:
We discussed the Q3 budget with Sarah and Tom. Tom flagged that hiring is running 20% over. Sarah suggested a freeze until October. No decision yet.

OUTPUT:
**Attendees:** Sarah, Tom
**Key issue:** Hiring spend 20% over budget
**Proposed action:** Hiring freeze until October
**Decision:** Pending

Now summarise these notes the same way:

[your actual notes]

The second prompt produces the right structure on the first try, every time. You didn't need to describe the fields — you showed them.

When to reach for it / when not to

Reach for it when: - The output has a fixed structure (tables, report sections, API payloads). - You've already iterated twice with instructions and the format keeps drifting. - You're building a prompt that will run many times — the example pays for itself.

Skip it when: - You genuinely want open-ended creative output. An example can constrain Claude more than you intend. - The task is simple and unambiguous. Adding an example just adds tokens.

One caution: if your example contains specific details (names, numbers, opinions), Claude may echo those details back into the real output. Use clearly fictional placeholder data in your examples — "Acme Corp", "£1,000", "John Smith" — so there's no bleed-through.

Try it

Pick a prompt you use regularly that still needs editing after the first response. Write one worked example of exactly the output you want, using dummy data. Paste it into the prompt before your real input. Compare the result to what you normally get. The format should lock in without any further instruction.