How to train an AI to write in your voice
Generic AI drafts sound like everyone. Here is how voice training actually works, what to feed it, and why an editable voice profile beats a black box.
To make an AI write in your voice, you need three things: real samples of your best writing, a tool that extracts your patterns into a profile it consults on every draft, and the ability to read and edit that profile when a draft sounds off. Prompting ChatGPT with "write casually" gets you nowhere; a maintained voice profile gets you drafts you edit instead of rewrite. Here is how to do it.
Full disclosure: we make Eden, whose voice Custom AI works the way this guide recommends. The principles apply to any serious voice tool.
Why "write it in my style" prompts fail
Pasting one post into ChatGPT and asking for "more like this" gives the model almost nothing durable. It imitates surface features of one sample, forgets everything next session, and regresses to the AI-average voice: balanced sentences, tidy transitions, the words every model overuses. Your voice is not one sample. It is a stable set of habits across many samples: sentence rhythm, vocabulary you use, vocabulary you would never use, how you open, how you land an ending, how you format.
Capturing that takes analysis across samples, stored somewhere permanent, applied on every draft.
What to feed the training
Quality beats quantity, and this is the step most people get wrong:
- Three strong samples beat twenty mixed ones. Pick the pieces that sound most like the writer you want more of, not everything you ever published.
- Match the format you want out. If you want tweets, feed tweets. Newsletter voice and thread voice are different registers; if you write very differently in two places, build two voices rather than one mushy average.
- Your best, not your oldest. Voice drifts. Feed the last year, not the archive.
How it works in Eden
Eden's builder (Custom AI, then Build a Voice) takes your pasted writing or your published work, studies it, and writes out a voice profile: your rhythm, sentence patterns, vocabulary, formatting habits, and explicit avoid/prefer rules.
The part that matters most: the profile is a real document in your Library, not a hidden setting. You can open it, read what the AI believes about your style, and edit the line that is wrong. When a draft sounds off, you fix the belief instead of re-prompting forever. The voice Custom AI reads the document live, so an edit changes every draft from then on.
Because it is an ordinary document, one voice is reusable everywhere: attach the same profile to your newsletter writer, your thread drafter, and (for agencies) a separate sealed voice per client built from that client's published material. And every reply shows a context line listing what it drew on, so you know the draft came from your profile rather than the model's imagination.
The workflow that compounds
Voice training is half the loop. The other half is deciding what the draft should say, and that comes from research rather than the model's memory:
- Find the outlier posts in your niche and save the patterns worth borrowing.
- Hand a pattern plus a topic to your voice AI: "take this hook structure, write it about X, in my voice."
- Edit lightly, then schedule it in the same workspace.
- When a draft sounds wrong, fix the voice profile, not just the draft. The correction is permanent.
If you work inside Claude or ChatGPT, your Eden Custom AIs are reachable there too through the Eden MCP, including as slash commands.
Frequently asked questions
How many writing samples does an AI need to learn my voice?
Fewer than you think. Three to five strong samples in the target format produce a sharper profile than a bulk dump, because mixed material averages your registers together.
Can I see what the AI learned about my style?
In Eden, yes: the voice profile is a document you can open and edit, and edits change future drafts immediately. If a tool cannot show you what it learned, you cannot fix it when it is wrong.
Will AI-written posts sound like AI?
Generic ones do. A profile-grounded draft in your cadence, lightly edited by you, reads like you on a productive day. The tell is never "AI was involved"; it is "no specific human was involved."
Can I train a voice for someone else, like a client?
Yes. Build one Custom AI per client from their published material, each sealed with its own voice profile and knowledge, so nothing bleeds between accounts. This is standard practice for ghostwriters and agencies in Eden.
Train yours in about ten minutes: start free, open Custom AI, choose Build a Voice, and paste your three best pieces. Then read the profile it writes about you. That part alone is worth the ten minutes.