How AI Can Draft a Cold Message That Sounds Like You
· 3 min read
Tell an AI model to "write in a friendly, professional tone" and it produces exactly what that sentence sounds like: competent, forgettable, and identical to what it would write for anyone else who typed the same instruction. The problem isn't the model. It's that a description of a voice and an actual sample of one are not the same input, and only one of them works.
Describing a tone doesn't transfer it
Most outreach tools ask you to fill in a few adjectives, maybe pick a "formality" slider, and call that personalization. What actually comes back from a model given "warm but direct" is whatever "warm but direct" statistically means across the entire internet — not what you mean by it. Two business owners who'd both pick "warm but direct" from a dropdown write nothing alike.
What works instead: real examples, not a description
The more reliable approach skips the adjectives and gives the model actual before-and-after pairs: a real message someone sent you, and the real reply you wrote back. Enough of those, and a draft stops reading like "AI, instructed to sound friendly" and starts reading like a message you'd plausibly have sent, because the model is matching a pattern instead of following an instruction about one.
This only works with a real sample, not a handful of invented examples typed in to seed it — three made-up messages teach the model your idea of what you sound like, which is a worse source than your actual writing. A WhatsApp chat export, scrubbed of anything identifying, is a surprisingly good source: it's unscripted, it's long enough to have real patterns in it, and it's the exact channel the draft will often go out on.
What this changes in a draft
- Sentence length and rhythm stop being uniform. If you write in short bursts, drafts start doing that instead of producing tidy three-sentence paragraphs every time.
- Stock phrases disappear. "I hope this message finds you well" is a tell precisely because it's nobody's real opening line; a model matched against real examples won't reach for it because it won't have seen you use it.
- The greeting and sign-off match what you'd actually type, not a formal template neither of you would use face to face.
The limits
This makes a draft sound more like you. It doesn't make the content true — a model that's good at matching your voice can still write a confident sentence about something it has no evidence for. Those are separate problems with separate fixes; voice-matching only solves the first one.
How AyePin does this
AyePin's Style page learns from a real WhatsApp export: it parses incoming-message/reply pairs, strips anything that identifies a real person before storing anything, and keeps each pair only as an anonymized example plus a vector used to find similar ones later. When drafting a new message, it pulls the closest-matching real replies you've actually sent and writes from those, not from a prior description of your tone. No raw chat is ever kept — only the matched pairs survive the import.