Copilot Few-Shot Prompting: How to Provide Examples in a Prompt

Few-shot prompting is a technique where you include examples directly in your prompt to guide Copilot’s output format, style, or content. Instead of asking a single question, you show Copilot one or more complete input-output pairs before your actual request. This method helps Copilot understand the exact structure you need, reducing vague or incorrect responses. This article explains how few-shot prompting works in Copilot, provides step-by-step instructions for creating effective examples, and covers common mistakes to avoid.

What Is Few-Shot Prompting and Why Use It with Copilot

Few-shot prompting is a method from large language model research where you provide a small number of demonstration examples within the prompt. Copilot reads these examples and infers the pattern you want, then applies that pattern to your new input. This technique works because Copilot’s transformer architecture learns from the sequence of tokens it sees — examples create a short-term pattern that the model follows.

You should use few-shot prompting when Copilot’s zero-shot output does not match your expected format. For example, if you ask Copilot to summarize an email and it returns a bullet list but you need a single paragraph, adding two examples of paragraph summaries will correct the output. Few-shot prompting also helps with domain-specific tasks like code generation, data extraction, or report writing where the default style is too generic.

The key advantage is that you do not need to train or fine-tune Copilot. You control the output by controlling what the model sees in the prompt. This makes few-shot prompting a flexible and immediate tool for business users who need consistent results across many similar requests.

How Few-Shot Prompting Differs from Zero-Shot and One-Shot

Zero-shot prompting gives Copilot no examples. One-shot prompting gives one example. Few-shot prompting gives two or more examples. Each additional example reduces ambiguity. For structured output like JSON, CSV, or tables, two to three examples are usually enough. For creative tasks like writing product descriptions, three to five examples help Copilot match a specific brand voice.

Steps to Write a Few-Shot Prompt for Copilot

Follow these steps to create a few-shot prompt that produces reliable output in Copilot across Microsoft 365 apps like Word, Excel, and Teams.

1. Define the output format you need
   Decide the exact structure, length, and tone of the response. For example, a table with three columns, a 50-word paragraph, or a JSON object with specific keys. Write this down before creating examples.
2. Write two to five example input-output pairs
   Each pair must have a clear label like “Input:” and “Output:” or use a consistent separator. Keep examples short — one sentence per input and one sentence per output. Use realistic data that mirrors your actual request.
3. Place all examples before your real request
   Start the prompt with the examples, then add a blank line, then write your actual input. Do not mix examples with the real request. Copilot treats the last input as the one to act on.
4. Use the same formatting in every example
   If you use bold labels, colons, or line breaks in one example, repeat the same pattern in all examples. Inconsistent formatting confuses Copilot and weakens the pattern.
5. Test the prompt and adjust example count
   Run the prompt once. If the output format is wrong, add one more example. If the output is too long, shorten the examples. Remove examples if the prompt becomes too long — Copilot has a context window limit of roughly 8,000 tokens in most Microsoft 365 apps.

Example: Few-Shot Prompt for Email Summarization in Copilot

This example shows a prompt that asks Copilot to summarize emails as one-sentence bullet points with a priority tag.

Input: “Team meeting rescheduled to Thursday 3 PM. New agenda attached. Please confirm availability.”
Output: “Team meeting moved to Thursday 3 PM — priority: high”

Input: “Office holiday party is December 15. RSVP by December 1. Potluck sign-up sheet in the break room.”
Output: “Holiday party December 15, RSVP by Dec 1 — priority: low”

Input: “Server outage reported in Chicago region. Estimated fix time 4 hours. All hands on deck.”
Output: “Server outage Chicago region, fix in 4 hours — priority: urgent”

After these three examples, paste the email you want summarized. Copilot will follow the same pattern.

Common Mistakes and Things to Avoid with Few-Shot Prompts

Copilot ignores the examples and writes in a different format

This usually happens when the examples are too long or the pattern is not obvious. Shorten each example to a single line. Use the exact same labels and punctuation in every pair. If Copilot still ignores the format, add a final instruction after the examples: “Follow the format above exactly.”

Copilot repeats the examples instead of processing the new input

This happens when the prompt does not clearly separate examples from the real request. Always put a blank line between the last example and the new input. Do not label the new input with “Input:” if you used that label in the examples — use a different label like “New:” or write the input without a label.

Copilot runs out of context and truncates the output

Each example consumes tokens. If you use five long examples, the remaining context may be too small for Copilot to generate a complete response. Limit examples to three for most tasks. Keep each input-output pair under 50 words. If you need more examples, split the task into multiple prompts.

Copilot produces inconsistent results across the same prompt

Copilot is non-deterministic — the same prompt can yield slightly different outputs. To reduce variation, use more examples. Four examples produce more consistent results than two. Also, specify the exact number of items in the output, for example, “Output exactly three bullet points.”

Few-Shot vs Zero-Shot vs One-Shot Prompting: When to Use Each

Item	Zero-Shot	One-Shot	Few-Shot
Examples provided	None	One	Two to five
Best for	Simple questions, common formats	Format guidance with one reference	Complex or strict format requirements
Output consistency	Low	Medium	High
Context window usage	Minimal	Low	Moderate
Example task	“Summarize this email”	“Summarize like this example”	“Summarize using these three examples as a guide”

Few-shot prompting is the most reliable method for structured outputs but uses more tokens. Use zero-shot for quick drafts and one-shot when you have a single good reference. Reserve few-shot for tasks where format accuracy is critical, such as data extraction, report generation, or code writing.

You can now build few-shot prompts in Copilot to get consistent, formatted responses for your business tasks. Start with two examples and increase only if the output does not match. For complex multi-step tasks, consider chaining few-shot prompts — use the output of one prompt as an example in the next. This technique works especially well in Copilot Chat in Microsoft Teams and Copilot in Excel for data transformation.