Most advice on matching AI voice stays abstract. It tells you to be specific without telling you what to type next. This is the opposite: a sequence you can run start to finish in one sitting and end with output that sounds like the target rather than the model's default register.
The process has six moves, performed in order. Each produces something concrete that feeds the next: a set of samples, a list of traits, a written rule block, a draft, a correction pass, and a final check. You do not need to understand the theory to follow it, though understanding helps when something goes sideways. Treat this as a recipe you can adapt once you have run it a few times.
If you want the conceptual background behind why these steps work, that lives in Making an AI Sound Like You Actually Wrote It. Here we keep moving.
Move One: Pull Three to Six Samples
Open a document and collect three to six pieces that everyone agrees sound right.
Pick the Right Ones
- Recent over old, so you capture the current voice.
- Same format as your target output: emails if you are writing emails.
- Strong over plentiful. Six great samples beat twenty mediocre ones.
Paste them into a scratch file. You will mine these for traits in the next move and reuse one or two directly in your prompt later.
Move Two: Extract Observable Traits
Read the samples like an editor and write down what they actually do, not how they feel.
List Specific, Checkable Habits
For each sample, note sentence length, whether contractions appear, vocabulary level, how it opens, and any signature punctuation such as em dashes or one-line paragraphs. You are turning a vague impression into a list someone else could imitate. If you find this hard, the gentler version is in Teaching an AI to Write in a Voice It Has Never Heard.
Keep the Traits That Repeat
Across your samples, some habits show up every time and some are one-offs. Keep the consistent ones; those define the voice. Drop the outliers.
Move Three: Write the Rule Block
Convert your trait list into instructions the model can follow.
Turn Adjectives Into Behaviors
Do not write "casual and confident." Write the behaviors that produce that effect: "Use contractions. Keep most sentences under twenty words. Open with a concrete situation. Avoid hedging words like maybe and perhaps." Every line should be something you could later check against the output.
Add One or Two Examples
Below your rules, paste one short excerpt from your best sample. The model anchors on the real text while the rules tell it what to notice. Rules plus example consistently beats either alone.
Move Four: Generate in Focused Chunks
Now produce a draft, but do not ask for everything at once.
Constrain the Task
A model handling a big task and a demanding voice will drop the voice to finish the job. Request one section at a time, so the model has attention to spare for style. For a long article, generate section by section rather than as a single block.
Restate the Voice Each Time
If you are generating across multiple turns, repeat the core rules with each request. Instructions fade as the conversation grows, and restating them keeps the voice steady.
Move Five: Correct With Targeted Edits
Your first draft will be close, not perfect. Fix it precisely.
Name the Exact Deviation
When a paragraph is off, say what is wrong and what right looks like: "This is too formal, rewrite with shorter sentences and at least one contraction." A specific correction teaches the boundary; "make it better" does not.
Do Not Regenerate From Scratch
Starting over throws away the parts that worked and reintroduces randomness. Edit what is wrong and keep what is right. This habit alone separates people who get consistent results from people who gamble.
Move Six: Check Against the Source
The final move is verification, and skipping it is the most common reason published AI text sounds off.
Compare Side by Side
Put the final draft beside a real sample and look for the traits from Move Two. Sentence length, contractions, opening style. Matching against the source beats trusting your gut, which drifts the longer you read. The patterns that slip past this check are gathered in 7 Common Mistakes with Prompting for Tone and Style Matching (and How to Avoid Them).
Inspect the Ending
Models drift back toward generic near the end of a long piece. Read the closing paragraphs specifically and fix any slide before you ship.
Adapting the Sequence to Your Work
The six moves are a complete recipe, but you will run them differently depending on what you are producing and how often.
One-Off Pieces Versus Ongoing Programs
For a single piece you may never repeat, you can run Moves One through Three quickly and informally, holding the traits in your head rather than writing a formal profile. For ongoing work, invest in writing the rule block down as a reusable profile. The few minutes spent formalizing it pay back across every future generation, and they protect you when a different person picks up the same voice.
Short Copy Versus Long Form
Short copy compresses the whole sequence. You still pull a sample and extract traits, but generation is a single short request and drift is a non-issue, so Moves Four and Six shrink to almost nothing. Long form stretches the opposite way: sectioned generation in Move Four and ending inspection in Move Six become the heaviest parts of the process because length is exactly where voice decays.
When to Slow Down
If the same deviation keeps reappearing across drafts, stop correcting individual pieces and go back to Move Three. A recurring problem means your rule block is missing a behavior, and adding one explicit rule fixes it everywhere at once. Repeated correction of the same flaw is a signal that the upstream encoding, not the output, needs work.
A Worked Micro-Example
To make the sequence concrete, picture matching a voice that is brisk and plainspoken.
From Sample to Rules
You pull three emails, notice they all run short sentences, use contractions, open with the point rather than a preamble, and avoid jargon. Those four repeating traits become your rule block. You paste one email as an anchor and state the four rules beneath it.
From Draft to Final
You generate a reply, find the opening too soft, and correct just that line: "Open with the answer, not a greeting." The rest already matches, so you leave it. A side-by-side with the original emails confirms the sentence length and directness line up, and you ship. That entire loop, sample to shipped, is the six moves at small scale.
Frequently Asked Questions
What if I only have one good writing sample to work from?
You can start with one, but pull more if you can. With a single sample you risk copying quirks that are not really part of the voice. Two or three let you spot which habits repeat and therefore actually define the style worth matching.
How detailed should the rule block be?
Detailed enough that every line is checkable against output, but not so long it buries the model. Five to ten concrete behavior rules plus one short example is a strong starting point. Add rules only when you see a specific deviation the existing ones do not cover.
Should I generate the whole piece at once or in sections?
In sections for anything longer than a few paragraphs. A model splitting attention between a large task and a demanding voice tends to sacrifice the voice. Smaller chunks keep the style tight and make corrections easier to apply.
Why correct instead of regenerating?
Regenerating discards the parts that already matched and rolls the dice again. Targeted correction preserves what works and teaches the model the specific boundary it missed, which converges on your target far faster than repeated full rewrites.
How do I know when the match is good enough?
When a side-by-side comparison with a real sample shows your key traits present throughout, including the closing paragraphs. If sentence length, word choice, and openings line up and a careful reader would not flag the voice as off, you are done.
Key Takeaways
- Run the six moves in order: pull samples, extract traits, write rules, generate in chunks, correct, verify.
- Convert mood adjectives into checkable behaviors and pair them with one short reference example.
- Generate in focused sections and restate the voice rules to keep style from fading across a long piece.
- Correct with specific, named edits instead of regenerating from scratch.
- Always verify against a real sample, paying special attention to the ending where drift concentrates.