The biggest complaint about AI writing is that it sounds like AI. The corporate blandness, the filler transitions, the way every paragraph feels algorithmically generated. Because it is.
But it does not have to be. You can train an LLM to write in your specific voice — your sentence rhythm, your vocabulary choices, your opinions. Not perfectly, but close enough that you spend five minutes editing instead of fifty.
This guide covers the exact process we use at Like One to maintain a consistent voice across all our content — part of the AI stack that runs our company, even though an AI writes the first draft of most of it.
Why AI Defaults Sound Generic
LLMs are trained on the internet. The internet is mostly corporate blog posts, Wikipedia articles, and content marketing. So the model’s default voice is a weighted average of all that — competent, informative, and utterly forgettable.
Your writing style is the opposite of an average. It is the specific choices that make your writing recognizable: short sentences or long ones, formal or casual, heavy on metaphor or straight to the point. The model knows how to produce all of these styles. It just defaults to the average unless you tell it otherwise.
Step 1: Collect Your Best Writing Samples
Find 5-10 pieces of writing that represent your authentic voice. Not your most polished corporate writing — your most YOU writing. Blog posts, emails, social media threads, journal entries. Anything where you were writing naturally, not performing.
What to look for:
- Pieces where people said "this sounds like you"
- Writing you produced quickly without overthinking
- Content that got engagement because of voice, not just topic
- Informal writing (emails, messages) — often more authentic than published work
Avoid: press releases, academic papers, or anything ghost-written. You want YOUR voice, not a professional veneer over it.
Step 2: Analyze Your Voice
Read your samples and identify the patterns. Look for:
Sentence Structure
Do you write short, punchy sentences? Long, flowing ones with multiple clauses? A mix where short sentences punctuate longer paragraphs? This rhythm is the heartbeat of your voice.
Vocabulary Level
Do you use simple words or SAT words? Industry jargon or plain language? Slang and contractions or formal phrasing? Note specific words you overuse — they are part of your signature.
Tone and Attitude
Are you earnest or sarcastic? Authoritative or conversational? Optimistic or skeptical? Do you hedge ("it seems like") or assert ("this is")? Your default emotional register defines your voice more than any other element.
Structural Habits
Do you start with a story or a statement? Use bullet points or prose? End with a call to action or an open question? These patterns are muscle memory that readers associate with you.
Step 3: Write Your Voice Profile
Distill your analysis into a concise voice profile. This becomes your custom instruction for the AI. Here is the format:
VOICE PROFILE:
- Sentence style: [short/long/mixed]. [specific pattern].
- Vocabulary: [simple/technical/mixed]. Uses [specific words/phrases].
- Tone: [casual/formal/mixed]. [specific attitude].
- Structure: [how you organize ideas].
- NEVER: [things that sound wrong in your voice].
- ALWAYS: [things that define your voice].
EXAMPLES OF MY VOICE:
[paste 2-3 short paragraphs that capture your style]Here is a real example:
VOICE PROFILE:
- Short sentences. Fragments occasionally. Then a longer sentence
that builds on what came before.
- Simple words. No corporate jargon. Uses "actually," "genuinely,"
and "the thing is" frequently.
- Direct and opinionated. States positions without hedging.
Occasional humor, never sarcasm.
- Opens with a bold claim or observation. Supports with specifics.
Ends with actionable advice.
- NEVER: "It is worth noting," "In today’s world," "Let’s dive in"
- ALWAYS: Uses contractions. Addresses reader as "you."
Includes concrete numbers when possible.Step 4: Set Up Your AI
In Claude
Go to your profile settings or create a Project. Paste your voice profile into the custom instructions. Upload 3-5 of your best writing samples as project knowledge. Tell Claude: "Match the voice and style of the uploaded writing samples. Follow the voice profile in your instructions."
In ChatGPT
Go to Settings > Personalization > Custom Instructions. Paste your voice profile in the "How would you like ChatGPT to respond?" section. You cannot upload documents here, but you can paste sample paragraphs directly into the instructions.
In the API
Include your voice profile in the system prompt. For production applications, include 2-3 example outputs as few-shot examples in the conversation history. This is the most reliable method because the model sees both the rules and the examples.
Step 5: Calibrate and Iterate
Your first outputs will be close but not right. This is normal. The calibration process:
- Generate a piece of content
- Read it out loud. Does it sound like you?
- Mark specific phrases that sound wrong
- Add those phrases to your NEVER list
- Mark things you would have said differently
- Add those patterns to your ALWAYS list
- Regenerate and compare
After 3-4 iterations, your voice profile will be tight enough that outputs need minimal editing. The key insight: the NEVER list is more valuable than the ALWAYS list. Preventing wrong-sounding phrases has more impact than requesting right-sounding ones.
Common Mistakes That Kill Your Voice
Most people fail at voice training not because the technique is wrong, but because they make one of these errors:
Mistake 1: Using Published Writing as Your Only Source
Published writing has been edited. Sometimes by you, sometimes by an editor, sometimes by both. It represents your polished voice, not your natural one. The AI needs your natural patterns — the ones that appear in first drafts, emails, and late-night messages. Mix published and unpublished samples for the most authentic profile.
Mistake 2: Over-Specifying Rules
A voice profile with 40 rules produces stilted, over-constrained output. The model spends so much effort following rules that it loses flow. Keep your profile to 10-15 core rules. The model will interpolate the rest from your examples. If your examples are good, less instruction produces better results.
Mistake 3: Copying Someone Else's Voice
You admire how someone else writes and try to train the AI on their style instead of yours. This always fails because you cannot edit convincingly in a voice that is not your own. When you read the output, you will not know if it sounds right because you do not have the intuition for someone else's patterns. Train on YOUR writing. Develop YOUR voice. Let influences show up naturally.
Mistake 4: Ignoring Context Shifts
Your LinkedIn post voice is not your newsletter voice. Your customer email voice is not your team Slack voice. One voice profile for everything produces content that sounds slightly wrong everywhere. Create 2-3 context-specific profiles instead of one universal profile. The investment pays for itself immediately.
Mistake 5: Never Updating the Profile
Your voice changes. The way you wrote two years ago is not how you write today. If your voice profile is based on old samples, the AI produces output that sounds like past-you — and you will not be able to pinpoint why it feels off. Refresh your samples and rules quarterly.
Measuring Whether It Actually Works
Voice training is subjective, but you can still measure progress:
Custom AI Training
Want AI that writes in your brand voice? Our consulting team builds custom training pipelines and style-matched AI systems for content teams.
- Edit time per piece: Track how long you spend editing AI drafts. Before voice training, expect 30-60 minutes per 1,000 words. After good training, expect 5-15 minutes. If editing time is not dropping, your profile needs work.
- The blind test: Show someone who knows your writing two paragraphs — one you wrote, one the AI wrote with your profile. If they cannot tell the difference (or guess wrong), you have succeeded.
- The wince count: Read the AI output and count phrases that make you wince. A new voice profile might produce 10-15 winces per page. A mature one should produce 1-2. Each wince is a rule you have not added yet.
- Consistency across pieces: Generate three different articles with the same profile. Read them back to back. Do they sound like the same person wrote them? Consistency is the hallmark of a well-trained voice model.
The ultimate test is whether you would put your name on the output without embarrassment. Not whether it is perfect — whether it is recognizably yours.
Advanced Techniques
The A/B Test
Generate the same content with and without your voice profile. If you cannot tell which one is yours, your profile needs work. The difference should be obvious.
Multiple Voices
You write differently in different contexts. Your blog voice is not your email voice is not your social media voice. Create separate voice profiles for each context. In Claude, use different Projects. (Not sure which model to use? See our ChatGPT vs Claude vs Gemini comparison.)
Voice Maintenance
Your writing style evolves. Update your voice profile every few months with recent writing samples. The profile that worked six months ago may not capture how you write today.
The Editing Pass
Even with a perfect voice profile, do a final editing pass. Read the output as if someone else wrote it. Change anything that makes you wince. AI gets you 80-90% there. The last 10-20% is what makes it genuinely yours. For even better results, build a personal AI assistant with persistent memory so your voice profile improves over time.
Voice Training by Platform
Different AI platforms have different strengths for voice training. The approach matters as much as the content.
Claude Projects (Best for Long-Form)
Claude Projects let you upload full documents as knowledge files. This means the model sees your complete writing samples — not just snippets in a system prompt. The result is better voice matching for blog posts, articles, and newsletters. The project-level custom instructions persist across conversations, so you set up your voice once and it stays calibrated. For writers producing regular long-form content, this is the most effective setup available.
ChatGPT Custom GPTs (Best for Templates)
Custom GPTs excel when you need the same voice applied to a repeatable format — weekly emails, product descriptions, social media templates. The instructions field gives you room for both voice rules and structural templates. The limitation is that you cannot upload writing samples as knowledge files in the same way, so voice matching depends entirely on your written rules and pasted examples.
API with Few-Shot Examples (Best for Production)
If you are building an application that generates content in your voice, the API approach is the most reliable. Include your voice profile in the system prompt and add 2-3 complete input/output examples as conversation history. This gives the model both explicit rules and implicit patterns to follow. The tradeoff is cost — more tokens in context means higher per-request costs — but the consistency is worth it for production use. Our guide to building AI agents covers how to integrate voice profiles into automated workflows.
When Voice Training Fails
Sometimes the output sounds wrong no matter what you do. Here is what is actually happening:
If the output sounds formal when you are casual, your examples are probably from published work. Add emails and messages. If the output sounds choppy, you have too many rules about sentence length. Remove constraints and let the examples do the work. If the output sounds like a pastiche — recognizably yours but somehow wrong — the model is over-indexing on your most distinctive phrases. Add more ordinary examples alongside your most stylistic ones. The goal is your everyday voice, not a caricature of it.
What Actually Works vs. What Sounds Good
Telling AI to "write casually" does not work. It produces a corporate person’s idea of casual. Telling AI "use sentence fragments, start paragraphs with conjunctions, and never use transition phrases" works because it describes specific behaviors.
Saying "match my tone" does not work. Uploading five paragraphs of your actual writing and saying "match this exactly" works because the model can pattern-match against concrete examples.
Rules without examples produce generic output. Examples without rules produce inconsistent output. Both together produce your voice. Start with your samples, build your rules from them, and keep refining until the output sounds like something you would actually publish.