How to Teach Your AI Assistant to Write in Your Own Voice
You teach an AI assistant your cold email voice by feeding it real emails you've actually sent — not polished marketing copy — along with specific corrections when its drafts miss, fed back in as examples rather than one-time instructions. A single instruction like "write in a friendly, direct tone" produces generic output; concrete examples of your actual phrasing, and a loop where you correct and re-feed drafts, are what make the output sound like you.
Teaching an AI assistant your cold email voice takes real sent examples and a correction loop, not a tone instruction. Here is the method that actually works.
"Write like me" is not an instruction an assistant can act on directly, because it has no idea what "like me" means until you show it. Voice is a pattern in word choice, sentence length, and what you leave out — not a style you can name in a sentence.
Voice is pattern matching, not instruction following
Tell an assistant to write in a "casual, confident tone" and you'll get casual, confident, and generic — the house style of every other business that gave the same instruction. Your actual voice is more specific than that: the way you open a sentence, whether you use contractions, how long your sentences run before you make the point, whether you ever ask a question in the first line. None of that is describable in an adjective. It's recognizable in examples, and an assistant can match a pattern far better than it can follow an abstract description of one.
What to feed it
The raw material matters more than the instructions wrapped around it.
- Real emails you've sent to actual prospects, including the ones that got no reply — the failures carry as much voice information as the successes.
- Notes or transcripts from calls where you explained what you do in your own words, which often sound more like you than anything you've written.
- Drafts you've rejected, with a short note on exactly what was wrong — too formal, too long, doesn't sound like something you'd actually say out loud.
- A short list of words and phrases you never use, which is often more useful than a list of words you do use, because it rules out an entire register of language in one line.
A handful of real examples, five or six, tends to do more than a long written description of your tone. The assistant isn't looking for a rule to apply; it's looking for a pattern to continue.
The correction loop
The first draft from any of this material won't be right, and that's expected. What makes the difference is what happens next: you mark exactly what's off, feed that correction back as context rather than just rewriting it yourself, and ask for another draft. Over a handful of rounds the gap closes, because each correction narrows the pattern the assistant is matching against. Skipping this step and accepting the first draft is the most common reason "AI-written" cold email reads as AI-written — not because the tool can't match a voice, but because it was only ever shown one example and no feedback.
This loop also catches the things a one-time instruction can't anticipate. You might not know you have a habit of ending emails with a specific kind of question until you see a draft that's missing it and you can tell something's off, even before you can articulate what. That's useful signal. Feed it back in.
Why this needs to be written down somewhere permanent
If the voice examples and corrections live only in one conversation, you start over every time you open a new session or switch assistants. The more durable approach is to capture the examples, the banned phrases, and the corrections that mattered as a standing reference — a file the assistant reads before it drafts anything, rather than a memory that resets. That's the difference between an assistant that writes in your voice today and one that writes in your voice every time, regardless of which tool you're using or how long it's been since the last campaign.
A portable skill file is exactly this kind of reference: it holds the outreach method itself — how a sequence is structured, what a first line needs to do — alongside the voice material specific to your business, so the two travel together. The method doesn't change between campaigns. The voice examples are what make the output sound like you wrote it instead of like a template with your name on it.
What to take away
- Tone instructions produce generic output; real sent examples produce voice matching.
- Include rejected drafts and failed sends — they carry as much information as the good ones.
- A list of phrases you never use is often more useful than a list of phrases you like.
- Correct drafts explicitly and feed the correction back in rather than silently rewriting them yourself.
- Keep the examples and corrections in a standing reference file so the voice holds up across sessions and tools, not just within one conversation.