AI drafts are fast — but they often sound too polished and generic. Five practical steps — gather representative samples, extract a concise style guide, use ChatGPT’s custom instructions or a bespoke GPT, include few-shot examples and iterate — will help the model mirror your voice and pacing so drafts need less line-by-line editing.
Why bother teaching an LLM your voice?
AI drafts are fast but often sound overly polished and generic. Training a model to mirror your mannerisms saves time moving from idea to a publishable draft and reduces editing. The result is not a perfect clone but a partner that better matches your tone, sentence rhythm and structure.
Step 1 — Gather representative samples
Start with three to ten pieces that show the voice you want to reproduce. Useful sources include:
- Blogs and long posts (for pacing and structure)
- Newsletters and essays (for argument flow)
- Short social posts (for punch and brevity)
- Transcribed voice notes (for spoken rhythms)
A mix helps the model learn both short-form and long-form patterns. Avoid rough drafts or heavily edited client work that don’t reflect your natural style.
Step 2 — Extract a compact style guide
Paste the samples into the model and ask for an analysis across tone, sentence length, vocabulary, pacing and personality. Request a short, reusable style guide — a few bullet points that act as your checklist. Example items: conversational tone; mostly short sentences with occasional long reflective ones; plain vocabulary; pragmatic, solution-focused voice.
Step 3 — Use platform personalisation features
Apply ChatGPT’s custom instructions or build a dedicated GPT and paste your style guide into the relevant fields so the model defaults to your tone across sessions. Set a clear rule such as: always respond in my style as defined below. A dedicated GPT can bake in preferred structures and formatting rules so the bot starts primed for tasks.
Step 4 — Few-shot examples and embeddings
Include one or two short samples directly in the prompt before requesting new text — a technique called few-shot prompting. That conditions the model mid-conversation to mirror your phrasing. Some users also distil multiple samples into a compact representation of their voice and reference that in prompts, giving the model a rulebook plus live examples each time.
Step 5 — Iterate and refine
Expect to adjust outputs. Read the first draft and give targeted feedback like “make it warmer,” “shorten sentences,” or “avoid corporate phrasing.” Repeat until the voice lands — practical tests show iteration matters more than a single perfect prompt. Keep a master prompt you can paste anywhere; for example: Write in my personal style as defined below. Style = [insert your style guide].
Related Articles
- Ubuntu to add two AI feature tiers over the next year
- Big Tech earnings collide with Powell on Wednesday
- Cohere and Aleph Alpha propose $20bn merger
Keep a reusable master prompt you can paste anywhere — for example: 'Write in my personal style as defined below. Style = [insert your style guide].' Also check your chosen model’s documentation for current context-window limits and supported personalisation features.
This article was created with AI assistance.