How to train AI to write like you
A useful writing clone is built from representative examples, explicit boundaries, and repeated evaluation—not a prompt that says “sound like me.”
Start with the job, not the model
Decide what the assistant should write before collecting examples. A style that works for client updates may not fit recruiting, clinical administration, or investor email. List the message categories, typical recipients, desired outcome, and situations that always require review.
Keep factual knowledge separate from writing style. Style describes how you communicate; knowledge determines what the message can safely claim. Mixing them makes both harder to maintain.
Choose representative writing samples
Use messages you actually sent and still consider good. Include several examples from each target situation: short acknowledgements, detailed explanations, declines, introductions, follow-ups, and difficult conversations.
Remove signatures, quoted threads, confidential identifiers, and text written by someone else. Ten varied, clean examples are more useful than hundreds of noisy exports. Add more only when they cover a missing situation.
Extract observable style traits
Avoid vague labels such as professional or friendly. Record traits that can be tested: average sentence length, greeting choice, use of contractions, paragraph length, directness, warmth, preferred sign-off, formatting, and phrases to avoid.
Separate consistent identity from situational adaptation. You might be concise with colleagues, explanatory with customers, and formal with counsel while still sounding like the same person.
Build an evaluation set
Hold back messages the system has not seen. Give it the same context and compare its draft with what you actually wrote. Score tone, structure, factual restraint, recipient fit, and the amount of editing required.
A draft that merely contains your favorite phrases is not necessarily faithful. The strongest measure is whether you would send it after a quick factual review.
Use edits as controlled feedback
Treat meaningful edits as signals, but do not learn blindly from every change. A rushed reply, one-off legal revision, or typo correction should not redefine your normal style.
Review the profile periodically. Remove accidental patterns, merge duplicate rules, and add examples for situations where drafts remain weak. Personalization is an ongoing calibration process.
Protect accuracy and relationships
Writing in your voice can make incorrect information more persuasive. Require human approval for financial, legal, clinical, personnel, contractual, and emotionally sensitive messages.
The objective is not to fool a detector. It is to reduce drafting time while preserving your intent, judgment, and responsibility for the final message.