Quick answer
Few-shot prompting means giving AI two to five short examples of your own writing before asking for the real task. The AI copies the pattern it can see — your sentence length, vocabulary and structure — instead of guessing from a description. It works on free plans and needs no technical setup.
Few-shot prompting is the fastest fix for AI writing that sounds nothing like you. Instead of describing your style in adjectives, you paste in two or three samples of your own writing and let the AI copy the pattern it can actually see.
Most professionals try the other route first. They write a careful paragraph explaining that they are direct, warm, and allergic to corporate filler — and get back something that reads like a press release with the corners sanded off. Then they spend fifteen minutes rewriting it, which was the exact work the AI was supposed to absorb.
✕ What a style description gets you
Hi team — wanted to circle back and share a quick update on where we've landed. We're thrilled to report that, following extensive cross-functional collaboration, the launch timeline has been revisited and realigned.
✓ What two writing samples get you
Team — launch moves to Thursday. We're finishing the database checks first. No customer impact. Timeline attached.
Same notes. Same AI. The only difference is that the second one had examples to work from. This guide covers what few-shot prompting is, the three-part structure that makes it reliable, and four prompts you can paste into ChatGPT, Claude, Gemini, or Copilot today — no code, no settings, and nothing that needs a paid plan.
Will this work for you?
What Few-Shot Prompting Actually Means
Few-shot prompting means giving an AI a small number of worked examples before you ask it to do the real task. The examples demonstrate the pattern you want — the tone, the sentence length, the structure, the things you deliberately leave out. The model matches that pattern instead of inventing one from scratch.
The word shot is just a slightly odd way of saying example, and it is the only piece of jargon this technique requires. Zero-shot is an instruction with no examples attached. One-shot gives one. Few-shot gives a handful. That is the entire vocabulary of few-shot prompting, which is a relief if you have ever bounced off a guide to prompt engineering written for software teams.
| Approach | What you give it | What it controls |
|---|---|---|
| Zero-shot | An instruction and nothing else | Topic and rough length |
| One-shot | A single worked example | Basic shape and format |
| Few-shot | Two to five varied examples | Voice, rhythm, structure, format |
Each example you add hands the AI more of the decisions you would otherwise have to fix by hand.
Why Style Descriptions Keep Failing
Describing your voice is lossy. “Direct but warm” has a few thousand valid readings, and a model handed that phrase will pick the most statistically average one — which is the tone you already dislike. Two real emails you sent last month have exactly one reading.
✕ Telling it your style
Write this in my voice: direct, warm, professional, concise, confident, and not too formal.
✓ Showing it your style
Match the tone and sentence length of the two emails below. Do not reuse their facts or topics.
What the AI picks up from a sample is far more specific than anything you would think to write down:
- How long your sentences run, and how often you break the rhythm
- Which words you reach for and which ones you never use
- Where you put the main point — first line or last
- What you leave out entirely, like greetings or sign-offs
This is the same principle behind giving AI proper context: concrete inputs beat abstract descriptions every time. OpenAI’s prompt engineering guide frames it as steering the model with a handful of input-and-output pairs rather than retraining it, and recommends varying those pairs so the pattern the model picks up is the one you meant. Few-shot prompting sits inside the broader shift toward context engineering — managing what the AI knows rather than polishing how you phrase the request.
The Sample-Fence-Target Method
The Sample-Fence-Target Method is the structure we teach at PromptPeakAI for making few-shot prompting hold up under real work: collect samples, fence them off so the AI copies the style and not the subject, then hand over the new task last. Three moves, in that order, every time.
The order matters more than the wording. Rules first, examples fenced in the middle, the new input last.
Collect two or three real samples
Open your sent folder and find writing you were happy with. Sixty to a hundred and twenty words each is plenty — a whole document is worse, not better, because long samples bury the pattern. Pick pieces on genuinely different topics. If both samples are about the same project, the AI will read that project as part of your style.
Fence the samples off
Wrap each sample in its own tag, like <example> and </example>, and add one blunt line telling the model to copy the tone and not the subject matter. Anthropic’s own prompting documentation recommends three to five relevant, varied examples wrapped in example tags for exactly this reason — the tags mark where the demonstration stops and the instruction resumes. The do-not line is a negative constraint, and it is the single highest-value sentence in the whole prompt.
Put the new task last
Every few-shot prompting template ends with the raw material for the actual job — your bullet points, the message you are replying to, the notes from the meeting. Nothing after it. The last thing the model reads is the thing it acts on, and anything you append below the target input competes with it.
You are writing a [document type] for [who reads it]. Match the tone, sentence length, and structure of the examples below. Do NOT reuse the facts, names, or subject matter from the examples. <example> [paste one piece of your own writing, 60 to 120 words] </example> <example> [paste a second piece, on a completely different topic] </example> Now write the [document type] using only this input: [your raw notes, bullet points, or the message you are replying to]
You should now have one reusable prompt with your own writing sitting inside it, ready to reuse tomorrow with a different set of notes.
A chat window with the master template pasted in, showing the rules at the top, two tagged writing samples in the middle, and the new notes at the bottom, just before pressing send.
Crop to the message composer only. Hide the sidebar, chat history, account name, and any profile photo.
File: few-shot-prompt-structure-in-chat.png · Alt: "A few-shot prompting template in a chat window with two writing samples wrapped in example tags"Four Prompts You Can Paste Today
These are the four workplace jobs where few-shot prompting pays back fastest. Each one follows the same three-part shape, so once you have adapted one you can adapt the rest in about a minute. If you write for a team rather than just yourself, the same samples become the seed of a shared voice everyone can produce.
Your weekly update to leadership
Status updates are the best place to start because you have dozens of past ones and they are short enough to paste whole. Swap my two examples below for two of your own — the point is the shape, not my wording.
You are writing my weekly project update for our leadership team. Match the tone, length, and structure of the examples below. Do NOT reuse the facts or projects named in them. Lead with the status. No opening pleasantries. <example> Launch moves to Thursday so we can finish the database checks. No customer impact. Timeline attached. </example> <example> Headcount budget is approved. Team leads, send hiring requests by 5pm Friday. No extensions this round. </example> Now write this week's update from these notes: [paste your raw bullet points]
A firm but friendly client reply
This is the one that saves the most agonising. The examples carry something you cannot easily describe: the exact ratio of warmth to boundary that works for your relationships. Use replies you were glad you sent.
You are writing my reply to a client. Match the warmth and the firmness of the examples below. Do NOT reuse their situations, numbers, or names. Hold the boundary, then offer a real next step. <example> Happy to get those live before your campaign. That work sits outside our current agreement, so I have put together a short addendum to cover it. Want me to send it over? </example> <example> I will pick this up first thing Monday and send a calendar invite for 9am so we can go through the numbers together. </example> Now write my reply to this message: [paste the client's message]
A post that sounds like you
Professional posts are where generic AI writing is most obvious, because your network has read your actual voice before. Feed it three of your own past posts and the difference is immediate — this pairs well with the structure in our guide to writing a LinkedIn post that earns engagement.
You are writing a short professional post in my voice. Match the sentence rhythm, line breaks, and vocabulary of the examples below. Do NOT reuse their topics. Do not use hype words or motivational filler. End on a plain question, not a call to action. <example> [paste one of your own past posts] </example> <example> [paste a second post on an unrelated topic] </example> Now write a post on this topic: [your topic in one line]
Messy notes into one format
Voice matching is not only about tone. The same technique standardises structure, which is why it works so well on the admin jobs nobody enjoys. Here one example is usually enough, because you want the format copied exactly rather than interpreted.
You are converting rough notes into our standard summary format. Copy the structure and the labels of the example below exactly. Do NOT carry over its content. If something is missing from my notes, write "not stated" rather than guessing. <example> INPUT: Met John re Q3 hiring. Two developers by September, one designer by October. Budget 180k. Sarah approved. OUTPUT: - Objective: Q3 headcount planning - Approved: 2 developers (September), 1 designer (October) - Budget: 180k, approved by Sarah </example> Now convert these notes: [paste your raw notes]
You should now have at least one of these four saved somewhere you can reach in five seconds, with your own samples already inside it.
A side-by-side view of a pasted writing sample and the AI output generated from it, showing the matched sentence length and structure.
Crop to the conversation thread. Hide the sidebar, account details, and any client or colleague names in the sample text.
File: few-shot-output-matches-sample-voice.png · Alt: "AI output next to the original writing sample showing matched tone and sentence structure"When Voice Matching Goes Wrong
Four failures account for almost everything that goes wrong with few-shot prompting, and each has a one-line fix. The most common by far is topic leaking: you paste a sample about a warehouse audit, and the AI cheerfully writes you a new paragraph about warehouses.
Topic leaking is a boundary problem, not an intelligence problem. Tags plus one do-not line solve it.
| What you see | Why it happens | The fix |
|---|---|---|
| It writes about your example topic | Nothing marked where the samples ended | Add tags plus one do-not line |
| Voice fades after two paragraphs | Your rules sat below the examples | Move the instructions to the top |
| Output arrives wrapped in stray tags | The model copied your own markup | End the prompt on a plain output label |
| It copies an odd word or a typo | Both samples share the same quirk | Swap one out for a more varied sample |
If the voice is still wrong after those four, the samples themselves are usually the problem. Writing you were lukewarm about produces output you will be lukewarm about, and the AI has no way to know which parts you were unhappy with.
What to Paste and What to Redact
Few-shot prompting runs on your real writing, which means your real work is going into a chat box. That deserves thirty seconds of thought before you paste, not a policy review after the fact.
Your own published posts, your own sent drafts, anything already public.
Client and colleague names, deal values, internal metrics, anything identifying.
Customer personal data, credentials, unreleased numbers, anything under NDA.
The redaction pass is quick and it barely costs you anything, because few-shot prompting teaches the AI your rhythm rather than your facts. Swap real names for role labels and round the numbers off — the voice survives intact.
On training, the boundary that matters is the type of account, not what you pay. Consumer accounts on the major tools generally do allow conversations to be used for model improvement, with an opt-out setting you have to find yourself, while business and enterprise accounts exclude that by default under contract. Our guide to using AI safely with work information walks through where those settings live.
Where Few-Shot Prompting Falls Short
Examples fix how something reads. They do nothing at all about whether it is true, and a matched voice makes a wrong claim more persuasive rather than less.
Three things stay yours no matter how good your samples are — these are the honest limits of the technique:
- Being right. Few-shot prompting cannot check a fact. It will match your cadence just as convincingly when the underlying claim is wrong.
- Judgement calls. Whether to send the firm version or the warm one is a read of a relationship, and no sample can carry that.
- Anything with legal weight. A contract clause or a formal warning needs a human behind the reasoning, not just the tone.
Two other limits are worth knowing. Very long prompts dilute the effect, so more samples stop helping somewhere around five and start hurting after that — quality and variety beat volume. And if you have read that you should adjust the temperature setting for better tone matching, that advice comes from developer tools: the consumer chat apps most professionals use do not expose it at all, so there is nothing to change.
People also ask whether they should be fine-tuning a model instead. For voice work, almost never. Fine-tuning wants hundreds of labelled examples and technical setup, while few-shot prompting wants two emails and a minute of your morning. Fine-tuning earns its keep on specialised, high-volume tasks — not on sounding like yourself in an inbox.
Stop Re-Pasting Your Samples Every Time
The reason people abandon few-shot prompting is friction, not results. Rebuilding the prompt every morning gets old by Thursday. The fix depends entirely on whether the job repeats.
One-off jobs get a pasted prompt. Anything you write weekly belongs in saved instructions instead.
For anything recurring, move the rules and the samples into whatever persistent slot your tool offers — custom instructions in ChatGPT, a project or a saved assistant elsewhere, or honestly just a note file you keep open. Which slot wins depends on how your tool ranks competing instructions, which our guide to system prompts and custom instructions unpacks properly.
Whichever route you take, remember that few-shot prompting is only as current as its samples — refresh them every few months. Your writing drifts, and a voice frozen in last year’s samples slowly stops being yours.
Build your voice prompt in ten minutes
- Pull two or three pieces of your own writing from your sent folder, 60 to 120 words each, on different topics. Write your rules at the top: what to produce, and one line saying not to reuse the examples' subject matter. Wrap each sample in its own example tags so the AI can see where the demonstration ends. Paste your new notes at the very bottom, with nothing after them. Save the whole thing somewhere you can reach in five seconds, and swap the samples out every few months.