Few-Shot Prompting: How to Make AI Sound Like Your Brand in Under 300 Words

AI is bad at sounding like you. Here's the only fix that works.

If you have ever asked ChatGPT or Claude to write a LinkedIn post "in my tone" and received something that reads like a 2019 corporate newsletter — generic, over-polite, weirdly optimistic — you are not alone. Out-of-the-box AI writing is flat because the model has no reference for what "your tone" actually means. "Be casual" and "write like me" are instructions the model cannot execute, because you have given it zero examples to pattern-match against.

There is exactly one prompting technique that fixes this reliably. It is called few-shot prompting, and it takes under 300 words of your own writing to make almost any modern model output content that sounds convincingly like your brand. This article shows you how to do it properly, where it breaks, and how to build a reusable few-shot block you can paste into any chat.

 

What is few-shot prompting?

Few-shot prompting is a technique where you give the AI model 2 to 5 complete examples of the exact output you want before you ask it to produce new output. The examples — called "shots" — show the model the structure, tone, length, and style you expect. The model pattern-matches against them and generates new content that follows the same pattern.

The term comes from machine learning literature. "Zero-shot" means you ask the model to do a task cold. "One-shot" means you give it one example. "Few-shot" means you give it a small set, typically two to five. Every mainstream AI tool — ChatGPT, Claude, Gemini — supports this out of the box. No plugins, no custom GPTs, no setup. Just examples in your prompt.

Stanford's IT training team calls few-shot prompting the single most effective technique for producing authentic brand content at scale. In practice, 2 to 5 examples is the sweet spot. More than five rarely improves output and just burns tokens.

 

Why does few-shot prompting work so much better than "write in my tone"?

Telling an AI "write in my tone" asks the model to guess from a vague label. Telling an AI "here are three examples of my tone — now write one more in the same style" gives it a concrete pattern to imitate. The model has no internal concept of your brand, but it is extremely good at pattern-matching when the pattern is visible.

This is the same reason junior copywriters improve faster when given a style guide plus three sample articles, rather than the style guide alone. The style guide tells you the rules. The samples show you what the rules actually look like on the page.

There is also a token-level reason. When you paste examples into a prompt, those examples enter the model's context window as raw text. The model gets to "read" your actual sentences — rhythm, word choice, punctuation habits, paragraph length — before it generates a single new word. That direct exposure is always more accurate than any adjective you could use to describe the style.

 

How many examples should you use, and of what?

Use three examples. Each one should be between 60 and 120 words of your own published writing — something that genuinely sounds like your brand at its best. Pick three pieces that are similar in format to what you want the AI to produce. If you want LinkedIn posts, show LinkedIn posts. If you want sales emails, show sales emails.

The quality of the examples matters far more than quantity. One carefully chosen post that nails your brand's rhythm will teach the model more than ten average posts. The mistake most people make is pasting in whatever happens to be at the top of their blog archive. Curate.

Also make sure the examples are diverse within your style. If all three examples are short and punchy, the AI will refuse to produce anything longer. If all three are 500-word think pieces, you will not get a snappy one-liner back. Show the range you want, within the voice you want.

 

How do you actually structure a few-shot prompt?

A good few-shot prompt has four parts: a role statement, the examples with clear delimiters, a clean task statement, and optional constraints. The delimiters matter more than most people think. Without them, the model confuses your examples with your instructions.

Use triple dashes (---) or XML-style tags (<example>) to separate each example. Claude in particular responds better to XML tags; ChatGPT and Gemini handle either. Put your examples in chronological or thematic order — not randomly — because models often weight later examples more heavily.

The clean task statement at the end should be a single, specific ask. Not "write me something similar" — that is vague. "Write a new LinkedIn post about our Q3 product launch, matching the voice and structure of the three examples above, between 80 and 120 words." That is specific enough to execute.

 

Where does few-shot prompting break down?

Few-shot prompting fails in four specific scenarios worth planning around. First: when your examples contradict each other — mixing a formal press-release tone with a casual Slack-style post confuses the model and you get mushy average output. Pick examples that are stylistically consistent.

Second: when the new task is structurally different from the examples. If your three examples are short social posts and you ask for a 2,000-word white paper in the same tone, the model has no scaffolding for the longer structure. Match example length roughly to target length.

Third: when your examples contain too many brand-specific proper nouns. The model will start hallucinating that those same proper nouns belong in the new output. Scrub or generalise client names, internal project codes, and dates unless you want them to reappear.

Fourth: when you keep the same few-shot block for months without refreshing. Your voice evolves. Update the examples once per quarter so the AI keeps matching your current output, not your output from last year.

 

What does a complete few-shot prompt look like?

Below is a full, copy-paste-ready template you can adapt for your own brand voice. Replace the placeholder examples with three of your own best pieces of writing. Keep the structure exactly as it is.

Try this prompt:

---

You are a senior copywriter for [Your Brand]. Your job is to match the voice and structure of the three examples below when generating new content. Do not invent new stylistic elements. Do not soften or sharpen the tone beyond what is shown.

<example_1>
[Paste your first piece of writing here — 60 to 120 words. Pick something confident and on-brand.]
</example_1>

<example_2>
[Paste your second piece of writing here — different topic, same voice.]
</example_2>

<example_3>
[Paste your third piece of writing here — show a third angle of your voice.]
</example_3>

Task: Write one new [LinkedIn post / email / blog intro] about [your topic], between [80 and 120] words, matching the voice, rhythm, and structure of the three examples above. Do not use any brand names, project codes, or proper nouns that do not appear in the examples. Output only the new piece — no preamble, no explanation.

---

This template works in ChatGPT, Claude, and Gemini with almost no modification. The only thing you should change between models is the delimiter style — Claude prefers XML tags (as shown), while ChatGPT and Gemini also accept triple-dash separators.

 

How do you turn this into a repeatable workflow?

Save your few-shot block as a text file on your desktop, named brand-voice-prompt.txt. Every time you need AI to write in your voice, open a new chat, paste the entire block, then add your new task. Total setup time: ten seconds.

If you use ChatGPT Projects, Claude Projects, or Gemini Gems, paste the few-shot block into the custom instructions for that Project. Every chat inside the Project will inherit the voice without you re-pasting. This is where few-shot prompting stops being a trick and starts being infrastructure.

Review the examples once per quarter. Rotate in any new piece of writing that you wish the AI would match, retire anything that now feels off-brand. Treat the few-shot block as a living document, the same way you treat a style guide.

 

What should you do in the next ten minutes?

Open the text editor of your choice. Find three pieces of your own writing that you are genuinely proud of — ideally from the last 90 days. Paste them into the template in the previous section. Save the file. Try it on one real task from your current backlog and compare the output to what you would have got with a generic "write in my tone" prompt.

The gap between those two outputs is the ceiling you have been hitting without realising it. Few-shot prompting lifts it immediately, and it costs nothing but the ten minutes it takes to curate the examples.

懂AI，更懂你 UD 相伴，AI 不冷。The practitioners who pull ahead in 2026 will be the ones who treat their own writing as training data — deliberately, quarterly, and for every repeatable output they care about.

 

Build Your Brand Voice Into Your AI Workflow

Few-shot prompting is the entry point. Building a full AI content system around your brand — with role prompts, review loops, and AI staff trained on your tone — is the next step. We'll walk you through every step, from prompt design to deployment.