"Write in a warm, conversational, authoritative tone."
Three adjectives, and every one of them is doing nothing. Not because they're wrong — because they're labels, and labels select from registers the model already has rather than transferring anything specific.
Show it three examples instead and you get something recognisably yours. That's few-shot prompting, and it's the highest-leverage technique in prompting for a reason worth understanding.
The short version
Zero-shot: instructions only. Fine for mechanical tasks.
Few-shot: two to five examples of what you want, then the task.
Why it works: examples encode specifics that adjectives can't — sentence length variation, how you open, whether you use fragments, where emphasis falls.
How many: three is the sweet spot for most marketing tasks. One gets copied too literally. Five or more starts averaging.
Why adjectives fail
Style words are compression, and the model decompresses them using its own defaults.
"Conversational" retrieves the model's conversational register. So does your competitor's "conversational," and everyone else's — which is precisely why that register has become so recognisable.
And adjectives are unfalsifiable. "Direct" doesn't specify sentence length, paragraph rhythm, whether you use contractions, or whether you'd ever write a one-word sentence.
Your voice is those specifics. It isn't a set of adjectives, which is why describing it doesn't transfer it.
What examples encode
Given three samples, a model can observe things you'd struggle to articulate:
- Average sentence length, and how much it varies
- Paragraph length, and whether you use one-liners for emphasis
- Whether you open with a statement, a question, or a scene
- Contraction use
- How direct your claims are
- Whether you hedge, and where
- Structural habits — how you signal a shift, how you close
None of that is expressible in adjectives, and all of it is observable in examples. That's the whole mechanism.
Where it works best
Voice and tone transfer. The obvious case. Three of your own pieces, and "match the rhythm, sentence length variation, and directness of these."
Format consistency. If you need output in a specific shape repeatedly — a particular brief structure, a report format, a standard email layout — one filled-in example transfers it better than a description of the fields.
Classification and extraction. Show three examples of input and correct output, and accuracy on the fourth improves markedly. Useful for tagging support tickets, categorising queries, extracting fields from messy text.
Judgement calls. Where "good" is hard to define but easy to recognise. Show three subject lines you'd send and three you'd reject, and the model gets closer to your standard than any description of your standard would.
That last pattern is underused. Negative examples — "here's what I'd reject, and why" — often transfer more than positive ones, because they mark the boundary.
How many, and how to pick them
Three is the practical sweet spot for marketing tasks.
One example gets copied. With a single sample, the model tends to mimic its specifics — the same opening move, the same structure, occasionally the same phrases.
Three shows the pattern and the variation. Enough to demonstrate what's consistent across your writing versus what changes per piece.
Five or more starts averaging. You're feeding it a wider range, and it produces the middle of that range — which is a step back toward generic.
How to choose:
Pick your best, not your most recent. You're setting a standard.
Match the format. Reference articles for a reference article. A LinkedIn post won't transfer usefully to a long-form guide.
Pick pieces that differ from each other in length and structure, so the consistent parts are your voice rather than one piece's shape.
And be explicit about what to match: "Match the rhythm, sentence length variation, paragraph length, and directness of these. Note that sections vary considerably in length and lists are not always three items."
Do examples use up context?
Yes, and it matters less than it used to.
Modern context windows are large enough that three articles as reference material is unremarkable for most tasks. It's worth knowing the constraint exists, without treating it as a reason to skip the technique.
Where it does become relevant: very long source documents plus long examples plus a long task. If you're pushing limits, shorten the examples to representative sections rather than dropping them.
One practical note: examples pasted near the start of a long prompt can matter less than the same examples placed close to the task. If your output isn't matching your samples, try moving them nearer the instruction.
Can you train AI on your old writing?
Two different things get called this, and the distinction matters.
Providing examples in a prompt — what this article describes. Nothing is learned or retained. Each conversation starts fresh, and you paste the examples again. Immediate, free, and completely under your control.
Fine-tuning — actually adjusting a model's weights on your material. Requires substantial data, technical setup, and cost, and it's overkill for a personal-brand site.
For almost everyone, few-shot prompting is the answer. It gets you most of the benefit with none of the setup, and it's easier to change — you can swap your reference set in seconds when your writing evolves.
One middle option: some platforms let you save persistent instructions or a custom assistant that carries reference material across conversations. That's still few-shot prompting, just stored rather than pasted. Setting that up.
What examples can't transfer
Worth being clear about the ceiling.
What transfers well: rhythm, sentence structure, formality, structural habits, format.
What transfers poorly: the judgement about what's worth saying, what to leave out, and which example illuminates a point. That reads like style from outside and is actually thinking.
What doesn't transfer at all: your opinions, your experience, your knowledge of the specific situation.
The implication: the more examples you give, the closer the surface gets — and the more the remaining gap is the substantive part. Which is why "make it sound like me" has a ceiling, and why the pieces worth publishing are the ones where you supplied something only you had. More on that.
Frequently asked questions
How many examples should I give?
Three, for most marketing tasks. One tends to get copied too literally — the model mimics its specific opening move and structure rather than the underlying pattern. Three shows both what's consistent across your writing and what varies per piece. Five or more starts averaging across a wider range, which moves the output back toward generic. Pick your best pieces rather than your most recent, and choose ones that differ from each other in length and structure.
Do examples use up context?
Yes, though modern context windows make three articles as reference material unremarkable for most tasks. It only becomes a real constraint when you're combining long source documents with long examples and a long task — in which case shorten the examples to representative sections rather than dropping them. One practical note: if output isn't matching your samples, try moving them closer to the task instruction rather than leaving them at the start of a long prompt.
Can I train AI on my old writing?
Two different things get called this. Providing examples in a prompt transfers style immediately with nothing retained between conversations — you paste them again each time, and it's free and fully under your control. Fine-tuning actually adjusts a model's weights, which needs substantial data, technical setup, and cost, and is overkill for a personal-brand site. Few-shot prompting gets most of the benefit with none of the setup, and it's easier to update as your writing changes.
What to do next
Pick three pieces of your own writing you're genuinely happy with — not your most recent, your best.
Save them somewhere you can paste from quickly.
That's your reference set, and it's reusable across every prompt you write from now on. Building it once is the highest-return five minutes in this whole category.
Free: The marketing prompt library.
Related guides
- Prompting for marketers — the wider picture
- The four-part prompt structure — where examples fit
- How to make AI writing sound like you — voice in practice
- Building a prompt library — storing your reference set
Join the Newsletter
Get practical marketing tactics delivered straight to your inbox.
Written by
Muhammad Basim
Related Articles
What Marketing Automation Actually Costs at Scale
Nobody's automation bill jumps because they added more automations. It jumps because they added steps to workflows that were already running — a filter here, an enrichment lookup there, a second notification — and the billing model charges for every one of them, on every run. That's the mechanic behind almost every "why did this […]
One Workflow, Three Tools: A Build Comparison
Feature tables don't tell you what a tool is like to use. The only thing that does is building the same thing twice and noticing where you got annoyed. So here's one realistic workflow — lead routing with enrichment and conditional assignment — specified once and mapped across Zapier, Make, and n8n. What changes between […]
GoHighLevel: 7 Automation Workflows to Build First
GoHighLevel gives you a hundred things you could automate, which is exactly why most accounts end up with forty half-built workflows and no measurable result. Seven are worth building first. They're the ones tied directly to revenue — catching leads before they cool, recovering appointments that would otherwise vanish, and asking for reviews at the […]