How to Make AI Write Like You

Pick samples that represent the voice you want to clone. Blog posts, emails, Slack messages, tweets - Whatever matches your target use case. 500-2000 words total is plenty.

If you're cloning a brand voice, pick recent published copy. If it's your personal voice, pick writing you're proud of, not just whatever's lying around.

Two selection mistakes ruin this step. First, mixing registers: your formal proposal and your group-chat messages average out into a voice that's neither. One voice per session. Second, including writing that was already heavily edited by someone else (or by AI) - You'd be cloning the editor. If you want to verify a sample reads as authentically you, run it through an AI detector first; anything that scores as machine-written is a bad teaching example.

Paste the examples and ask:

"Below are 3 samples of my writing. Analyse the voice in detail: typical sentence length, vocabulary level, use of contractions, attitude towards the reader, recurring phrases, what's distinctive. Be specific. Don't be flattering."

You'll get back a 200-400 word voice description. A good one reads like: "Short declarative sentences, average 12 words. Starts paragraphs with a claim, ends them with a jab. Heavy use of contractions and sentence fragments for emphasis. Addresses the reader as a smart equal. Allergic to corporate vocabulary. Recurring move: undercutting a serious point with a dry aside." If yours comes back with horoscope-grade filler ("confident yet approachable"), reply: "Too generic. Quote specific sentences from my samples as evidence for every claim."

The first analysis usually misses something. Push back:

• "You missed that I use semicolons heavily."
• "My tone isn't 'professional' - It's mildly sarcastic."
• "I avoid the word 'leverage' on principle. Add that."

A trick that surfaces blind spots: ask "Write one paragraph about remote work in this voice" and study what feels off in the result. Whatever feels wrong is a missing rule - Name it and add it. Most people need 2-3 rounds. You're done when the description doubles as a style guide you'd hand a human ghostwriter, including the negative rules ("never opens with a question," "no exclamation marks"). Negative rules do more work than positive ones; AI defaults are the enemy.

The final voice description is now your system prompt for any future writing task. Format like:

"You are writing in [Name]'s voice. Voice characteristics: [paste analysis]. Always match this voice. When in doubt, err toward [your dominant trait, e.g. 'blunt over polite']."

Save it in your AskAI.free saved prompts library with the Pro plan, or keep it in a notes file and paste it at the top of any chat. The tie-breaker line at the end matters more than it looks: when the model is unsure, it falls back to its bland defaults unless you've told it which way to lean.

Combine the voice prompt with a specific task: "Write a 200-word LinkedIn post about [topic] in this voice." The first output may still feel slightly off. Iterate:

• "More direct, less explanatory."
• "Add more humour - I'm not this serious."
• "Cut the hedging - Pick a position."

After 2-3 iterations, the AI should produce drafts that need only light editing to feel like you. Expected quality at this point: cadence and vocabulary land reliably; what's still missing is content only you have - Anecdotes, opinions, the story from last Tuesday. Feed those in as raw bullet points and let the voice prompt handle the phrasing. Voice cloning gets you the how; you still supply the what.

A voice prompt that only works on easy topics isn't done. Test it where AI defaults are strongest: ask for an apology email, a piece of bad news, and a topic you know nothing about. These pull hard toward corporate mush ("We sincerely apologise for any inconvenience..."), and a weak voice prompt collapses immediately.

If the voice survives an apology email - Still sounds like you, just contrite - It's robust. If not, you're usually missing emotional-register rules. Add lines like: "Even when apologising, stays plainspoken; never uses 'inconvenience', 'regret to inform', or passive voice to dodge blame." Then re-test. This step takes 5 minutes and is the difference between a party trick and a tool you trust on real correspondence.