How to Make AI Write in Your Voice (So It Stops Sounding Generic)
You know the feeling. You paste a prompt into ChatGPT, hit enter, and get back something that reads like the most confident, least interesting version of the internet. Technically correct. Structurally fine. But not you. Not even close.
Here's the thing: AI doesn't have a voice problem. It has a signal problem. When you give it a blank prompt, it defaults to statistically safe, high-frequency language — the kind that sounds vaguely like everything and exactly like nothing. The fix isn't a better model. It's teaching the model what you sound like, before you ask it to write anything. That's what this post is about. See also: why AI writing sounds robotic and how to fix it for the broader context.
Why AI Sounds Like Everyone — and No One
AI generates text by predicting the most likely next token given everything that came before it. Without any personal writing signal, "most likely" means "most common across all training data" — which produces polished, generic prose. It's not a flaw. It's the math doing exactly what the math does.
A 2023 Stanford HAI analysis of large language model outputs noted a consistent tendency toward high-frequency, statistically central language patterns when prompts lack personal or domain-specific context. In plain terms: vague input, average output. A 2024 Content Marketing Institute survey found that 58% of content marketers named "maintaining brand voice" as their top challenge when adopting AI writing tools — not quality, not speed, not accuracy. Voice.
The problem compounds because most people write prompts like search queries. "Write a LinkedIn post about productivity." That's 6 words of context for a system trained on billions of documents. It will write a LinkedIn post. It just won't write your LinkedIn post.
Generic prompt vs. voice-primed prompt — the same output task
| What the prompt contained | What the output sounded like | Why |
|---|---|---|
| Topic only ("write about X") | Smooth, structured, no personality | No voice signal — defaults to average internet tone |
| Topic + 3 style samples from your past writing | Recognizably similar sentence rhythm and word choice | AI pattern-matches your samples and replicates the style |
| Topic + samples + tone descriptors + banned words | Very close to your actual voice on first draft | Multiple converging signals narrow the output space |
Here's what that looks like in practice. Same task — a short intro for a blog post about time management — with and without a voice profile.
The right column isn't "better writing." It's a specific person's writing. That's the distinction that matters when you publish under your own name.
Your Voice Profile — The 4 Ingredients AI Needs
A voice profile has four components: 3–5 writing samples from your own work, a tone descriptor (3–5 adjectives plus a "sounds like X, not Y" contrast), a few-shot example pair showing generic-to-your-style rewriting, and a banned-word list. Together, they constrain the model's output toward your actual voice rather than the statistical center of the internet.
The 4-ingredient breakdown
| Ingredient | What it is | Example | Where it goes in the prompt |
|---|---|---|---|
| 1. Style samples | 3–5 paragraphs from your best work | A newsletter intro, a LinkedIn post, a blog section you're proud of | Top of prompt, before any task instruction |
| 2. Tone descriptor | 3–5 adjectives + one contrast statement | "Direct, slightly irreverent, skeptical of jargon. Sounds like a smart friend, not a consultant." | After samples, as a one-sentence paragraph |
| 3. Few-shot example pair | One "before" (generic) + one "after" (your rewrite) | Show the model exactly what transformation you want | After tone descriptor, labeled Before/After |
| 4. Banned-word list | Words and phrases you never use | "Never use: leverage, utilize, delve, unlock, game-changer, it's worth noting" | Last line before the task instruction |
Why all four? Style samples show the pattern. Tone descriptors label what makes the pattern distinctive. The few-shot pair demonstrates the transformation explicitly. The banned list removes the most persistent offenders. Each ingredient targets a different failure mode. Omit one and you'll still get better results than a bare topic — but you'll hit a ceiling.
The few-shot example pair is particularly powerful. Research on context learning in language models (including the original GPT-3 paper by Brown et al., 2020) showed that providing even one or two examples in-context dramatically narrows output distribution toward the demonstrated style. You're using the same mechanism — just aimed at your voice rather than a benchmark task.
Step-by-Step: Teaching AI to Write Like You
Paste your style samples first, then your tone descriptor, then the before/after example pair, then your banned-word list, then — and only then — your actual task. That order matters. Everything before the task instruction is signal. The task itself is only the last 10% of the prompt.
Pick paragraphs where you sound most like yourself — not your most polished work, your most you work. 100–200 words total is enough.
3–5 adjectives that describe your style. Add one "sounds like X, not Y" contrast to anchor it. ("Conversational and punchy. Sounds like a knowledgeable friend, not a thought leader.")
Take a generic AI sentence, rewrite it in your voice, and label both. This is your most powerful teaching signal.
The 5–10 words or phrases you most want to avoid. "Leverage," "utilize," "delve into," "it's worth noting," "game-changer" are common starting points.
For a deeper guide on prompt structure more broadly, the same principles apply — front-load context, be specific about format, label everything.
Here's how a full voice-primed prompt session looks:
(Role) You are a writing assistant trained to replicate a specific author's voice.
(Context) Here are 3 writing samples from the author. Study the sentence rhythm, word choice, and tone:
Sample 1: "Most productivity advice is optimized for people who already have their act together. The rest of us need something more forgiving."
Sample 2: "I don't think the problem is focus. I think the problem is that we're trying to focus on too many things at once, which is a different problem with a different solution."
Sample 3: "Here's what no one tells you about delegation: it gets harder before it gets easier. The first month you'll spend more time explaining than doing. That's the cost. It's worth paying."
Tone: Direct, slightly skeptical, conversational. Sounds like a smart colleague who respects your time, not a motivational speaker.
Before: "It is essential to prioritize tasks effectively to maximize productivity."
After: "Pick one thing. Do it first. Everything else is noise."
Never use: leverage, utilize, it's worth noting, game-changer, impactful.
(Task) Write a 3-sentence opening for a newsletter about inbox zero.
(Format) Plain prose, no bullet points, under 60 words.
Your inbox is not a to-do list. It never was. The reason inbox zero feels impossible isn't discipline — it's that you're managing a queue that other people are allowed to add to at any time, without asking.
That output didn't come from a better model. It came from a better prompt. The samples, tone descriptor, before/after pair, and banned-word list each contributed. Remove any one of them and you'll notice the difference.
5 copy-paste voice-profile prompt cards
1. Full Voice Profile (samples + tone + banned words)
2. Quick Tone-Only Version (no samples)
3. Few-Shot Rewrite (teach by example)
4. Banned-Word Filter Pass
5. Full System Prompt (for Custom Instructions / Projects)
Advanced: Locking Your Voice with a System Prompt
If your AI tool has a system prompt or custom instructions field — ChatGPT's "Custom Instructions," Claude's "Project Instructions," or a developer-configured system prompt — move your entire voice profile there. Every session then starts with your voice pre-loaded. No copy-pasting. No reminding. The model walks in already knowing how you write.
ChatGPT's "Custom Instructions" (available to Plus and Team subscribers) has two fields: "What would you like ChatGPT to know about you?" and "How would you like ChatGPT to respond?" The second field is your voice profile. Paste Prompt Card 5 there, customized to your samples and preferences.
Claude's Projects feature works similarly: each Project has an "Instructions" block that persists across all conversations in that project. If you do most of your writing work in one Claude Project, your voice profile lives there permanently.
[System prompt / Custom Instructions — already set in account settings]
You are a writing assistant for Alex. Write in Alex's voice — direct, skeptical of buzzwords, conversational. Samples on file. Never use: leverage, utilize, delve, game-changer, it's worth noting.
[System acknowledged — voice profile active for this session]
Write a 3-paragraph LinkedIn post about why most meeting agendas don't work.
Most meeting agendas are just a list of topics with a time budget nobody sticks to. That's not an agenda. That's a wish list.
A real agenda answers two questions: what decision are we making, and who needs to be in the room to make it. If you can't answer both, cancel the meeting. Send a Slack message instead.
The reason this is hard is that most agendas are written by whoever scheduled the meeting, not whoever has to run it. Fix that first.
What to store in the system prompt vs. per-message context
| Keep in system prompt (persistent) | Keep in per-message context (task-specific) |
|---|---|
| Voice samples (3–5) | Specific topic and audience for this piece |
| Tone descriptor | Desired word count and structure |
| Banned-word list | Source material, research, or data to include |
| General writing rules (sentence length, POV) | Examples relevant only to this specific piece |
One practical note: system prompts consume tokens. Very long voice profiles (500+ words of samples) can crowd out the conversation context on shorter-context models. Keep system prompt samples to 3 short excerpts — 60–80 words each — rather than pasting entire articles.
Frequently Asked Questions
How many writing samples do I need to give AI for it to copy my style?
Three is the practical minimum. Below three, the model doesn't have enough signal to distinguish your patterns from the general noise of the training data. Five samples is a good ceiling — above that, you're adding context tokens without meaningfully increasing voice accuracy. Pick samples that are representative, not exceptional. Your most "you" paragraphs, not your most polished ones.
Will this work with Claude, Gemini, and ChatGPT equally well?
The technique works with any instruction-following model. The degree of voice replication varies — more capable models (GPT-4o, Claude 3.5+, Gemini 1.5 Pro) show stronger stylistic adaptation from the same samples. The prompt structure is identical across tools. Test with one model, then reuse the same voice profile elsewhere without modification.
What kinds of writing samples work best?
Published paragraphs where you wrote naturally under pressure — newsletter intros, social posts, sections of blog posts you didn't over-edit. Avoid highly edited or co-written pieces where your voice got smoothed out, and avoid very short samples (single sentences) that don't contain enough rhythm to pattern-match. 50–100 words per sample is the sweet spot.
My output still sounds a bit off after using the voice profile. What am I missing?
Usually one of three things: the samples aren't representative of your actual everyday voice (too formal, too polished), the tone descriptor is too vague ("professional and friendly" describes half the internet), or the banned-word list is too short. Add one more specific sample, sharpen the contrast statement in your tone descriptor, and expand the banned list to 10 items. One full editing pass where you rewrite AI sentences into your own words also gives you new material for the few-shot example pair.
Is there a risk that training AI on my voice makes it too predictable?
The voice profile constrains style and tone — it doesn't constrain ideas, structure, or argument. The output still varies based on the task prompt. What you're removing is the random stylistic noise that makes AI output sound generic, not the variation that makes writing interesting. Think of it as keeping your accent consistent while still saying different things.
How often do I need to update my voice profile?
When your writing evolves meaningfully. Most writers' voices shift slowly — if your samples are from two years ago and you've changed significantly, add one or two recent samples and remove the oldest. If you write in different contexts (newsletter vs. LinkedIn vs. long-form blog), consider keeping a separate voice profile for each context. Voice shifts are real but gradual; annual reviews are usually sufficient.
The Takeaway
AI doesn't default to generic output because it's bad at writing. It defaults to generic because you haven't given it a reason to do otherwise. The four-ingredient voice profile — style samples, tone descriptor, few-shot example pair, banned-word list — is that reason. Build it once, store it in your system prompt, and every piece you generate starts with your voice already active.
The goal isn't AI that sounds exactly like you wrote it with zero edits. It's AI that produces a first draft close enough to your voice that revision takes 15 minutes instead of an hour. That's a different problem, and this is its solution.
Comments
Comments (0)
Leave a Comment