Getting an AI model to write in your own voice — rather than its polished, faintly corporate default — is one of the most useful prompting skills. The technique is called style transfer: you supply examples of a target voice and ask the model to produce new text that matches it. Done well, it turns a generic assistant into a convincing ghostwriter. Done badly, you get text that is technically on-topic but sounds like everyone else’s AI output.
Show, don’t describe
The single biggest lever is examples. Adjectives like “casual” or “punchy” mean different things to different models, but a real passage of your writing is unambiguous. Paste two to four representative samples and instruct the model: “Study the voice, rhythm, vocabulary, and punctuation of these passages, then write the following in the same style.” The model can then infer your average sentence length, your fondness for em dashes or short fragments, and your level of formality directly from evidence rather than guesswork.
Name the attributes that examples miss
Examples capture a lot, but some traits are worth stating explicitly. Useful attributes to specify include: sentence length and variation, formality level, use of contractions, first vs third person, humour or dryness, jargon tolerance, and rhetorical habits (rhetorical questions, lists, direct address). A short checklist appended to your samples — “uses contractions, avoids exclamation marks, favours concrete nouns over abstractions” — sharpens the imitation and stops the model from sliding back to its defaults.
Lock the voice with a system prompt
For anything longer than a single reply, move the voice rules into a system prompt. A system prompt persists across the whole conversation, so the model keeps the style without you re-pasting it every turn. A workable template is: “You are writing as [name/role]. Voice: [3–5 attributes]. Here is a sample of the target style: [one passage]. Match this voice in everything you write.” Then send ordinary task messages. This keeps long drafts consistent and frees the context window for the actual content.
Test, then correct with contrast
After the first output, do not just accept or reject it — diagnose it. Ask the model to compare its draft against your sample and list three differences, then revise. You can also use contrastive correction: “That is too formal and the sentences are too uniform; vary the length and loosen the tone.” Pointing at the specific gap between the output and the target is more effective than vague re-prompting, because it tells the model exactly which dimension to move.
Common failure modes
Three problems recur. First, regression to the mean: over a long session the model drifts back to its default register, so re-anchor periodically. Second, over-imitation, where the model copies the surface tics of your sample (specific phrases, topic) rather than the underlying voice — fix this by using varied samples and reminding it to imitate style, not content. Third, flattening, where genuinely distinctive quirks get smoothed away; counter it by explicitly protecting them (“keep the abrupt one-line paragraphs”). Treat style transfer as an iterative loop — sample, generate, compare, correct — and the AI will sound increasingly like you and less like itself.