What it is
One example of the output you want, pasted into the prompt with a sentence that says which of its features to copy: structure, length, tone, level of detail. On the ten checks this is Reference, the eighth check. A prompt with no example is called zero-shot; one example is one-shot; several are few-shot.
An example carries information that is tedious to describe. How long each line runs, whether there is a greeting, how numbers are written, how blunt the bad news is. A model reads all of that from one good sample and applies it to new content.
When it helps
It helps most with anything you produce again and again: the weekly update, the monthly minutes, replies to a common customer question, product listings. You already have an approved version, and matching it is exactly the job.
It helps when your house style is hard to put into words. “Our tone is friendly but direct” could mean many things. An email your team was happy with means one thing. And it helps with small fixed formats, such as one line per ticket with set fields, where a single example row settles the order and the separators at a glance.
When it does not
Models copy examples closely, sometimes too closely. With one sample, the new answer may borrow its phrases, its length to the word or even its facts. Say what to copy, such as “the structure and length, not the wording”, and keep the example in its own fenced block so last week’s numbers do not end up in this week’s update.
One example narrows the range. If inputs vary a lot, every answer will still look like the one sample, including where it should not. For varied work, use a few examples that differ from each other, or show a good one next to a weak one with examples and counterexamples.
An example shows a format, not the rules behind it. One sample ticket labeled “Login” does not tell the model what the other categories are or where the boundary between “Bug” and “Login” lies. Give the list, and the rule for borderline cases, as words.
A weak example gets you weak answers that match it. Use your best one. And check it for names, figures and other details that should not travel into new text; Safe Paste can swap them for placeholders first.
Before and after
Two requests where the after version adds one example. Each score comes from the Prompt Grader’s free ten-check estimate, written for any chat model.
Write this week's status update for my director from my notes. Notes: """ [this week's notes] """
Write this week's status update for my director from my notes. Match the structure and length of last week's update, below. Copy its shape, not its wording. Notes: """ [this week's notes] """ Last week's update: """ Shipped: invoice export for EU customers (two days late, now live). In progress: bulk discounts, 60% done, on track for the 14th. Blocked: tax rules for Canada, waiting on finance since Monday. Decision needed: whether to delay the pricing page until bulk discounts ship. """
Without the example, the model picks its own shape for a status update, which may be a paragraph per project. With last week’s update, it returns four labeled lines in the same order, so your director finds this week’s news where they found last week’s. The “shape, not wording” line keeps it from reusing last week’s phrases.
Sort these support tickets into categories. """ [the tickets, one per line] """
Sort these support tickets into categories. Use the format of this example, one line per ticket: App logs me out every hour | Login | Medium Categories: Login, Billing, Bug, Feature request, Other. Urgency: Low, Medium or High. """ [the tickets, one per line] """
Without an example, the categories can come back in any form, from a grouped list to a paragraph per ticket. With one example row and the two lists, every ticket comes back as one line with the same three fields, ready to paste into a spreadsheet. This pair gains four points because the example also sets a format and a size for each line.
The line to add
Match the {{features_to_copy}} of the example below, with new content. Copy its shape, not its wording or its facts.
Example:
"""
{{your_best_past_example}}
"""
Name the features to copy plainly: structure and length, tone, the order of sections, how much detail per item. If the answer still copies too much, the fix is usually a second example that differs from the first, not a stronger warning.
In ChatGPT, Claude, Gemini and Copilot
Every vendor recommends examples, and the ones that give a number suggest more than one. Anthropic’s guide for Claude calls examples one of the most reliable ways to steer format, tone and structure. It asks for examples that are relevant, varied enough that Claude does not pick up unintended patterns, and wrapped in <example> tags, and it suggests three to five for best results.
Google’s guide for Gemini recommends always including few-shot examples, warns that too many can make the model overfit its answers to them, and asks for the same structure and formatting across all of them. OpenAI’s prompt engineering guide suggests showing a diverse range of possible inputs with the desired outputs. Microsoft’s guide to Copilot prompts includes samples in what it calls the source, the information or samples Copilot should use.
One example is where to start in any of them. Add more when the answers copy the first one too closely or the inputs vary.
Key takeaways
- One good past example sets length, order and tone faster than a description.
- Say which features to copy and which to leave, such as wording and facts.
- Keep the example in its own fenced block so its facts stay out of the new answer.
- An example shows a format, not the rules. List categories and boundaries in words.
- When answers copy too closely, add a second example that differs from the first.
Questions
What is one-shot prompting?
Giving the model one example of the output you want, alongside the instructions. Zero-shot means no example and few-shot means several. One example is often enough to set format and tone, and it is the easiest to find, because most people have one good past version of whatever they need to produce.
How many examples should I give ChatGPT or Claude?
Start with one. If answers copy it too closely, or your inputs vary a lot, add a few more that differ from each other. Anthropic suggests three to five varied examples for best results, Google warns that too many can make answers overfit to them, and OpenAI suggests examples that cover a diverse range of inputs.
Why does AI copy my example word for word?
Because the example is the clearest signal in the prompt. Say which features to copy, such as structure, length and tone, and which not to, such as the wording and the facts. Put the example in its own fenced block, and if it still copies too closely, add a second example that differs from the first.