Guides ยท 22 November 2026
How to make AI write like you: we tested five ways
Most advice says to ask AI to read your writing and produce a style guide. We tested that against four other ways. In our test, pasting in two real pieces you wrote beat every written guide, and adding a guide on top made it worse.
The short answer. Give the AI two recent pieces you wrote, of the same kind as the one you want. Give it notes that hold everything you want to say. Keep one piece back and use it to check the result. Add a rule only when a check shows it helps.
Where this started
Michael Taylor, who leads AI testing at the publication Every, spent years trying to get AI to write like him. In November 2026 he published what finally worked. His point is that a style guide is a list of guesses about how you write until you test it against writing it has not seen.
Read the original: How I Finally Got AI to Write Like Me, by Michael Taylor at Every. The testing idea below is his. The test, the numbers and the conclusions on this page are ours.
His method, in brief: turn each of your published pieces into plain notes that keep the facts and drop the style. Have the AI write a draft from those notes. Then ask a second AI which of two drafts is closer to what you really wrote, twice, with the drafts swapped, because AI judges tend to favour whichever they read first. He used software that rewrites the guide automatically over many rounds, and reports that guides built this way beat a guide Claude simply wrote on 80 to 95 percent of the posts he held back.
That set-up needs code and a budget. We wanted to know what an ordinary person with a chat window should do. So we ran a smaller version ourselves.
What we tested
We took 12 published blog posts by one writer, between 880 and 4,100 words each. We held back the four most recent as the exam. The AI that wrote the drafts never saw them. For every post, an AI made plain notes: the argument, the outline and every fact, with none of the wording.
Then we had the same AI model (Claude Sonnet) write each exam post from its notes, five ways:
- No guide. Just "you are a ghostwriter for a founder who blogs about startups".
- A described guide. A stronger model read eight of the writer's posts and wrote a one-page style guide. This is what most people do.
- A tested guide. We took the described guide, wrote drafts with it, had a model list how each draft differed from the real post, and rewrote the guide from those notes. Twice.
- Two examples. No guide at all. We gave the AI two of the writer's recent posts and said "match this".
- Two examples plus the tested guide.
A separate AI judge (Claude Opus) saw the real post and two drafts with no labels, and picked the draft closer in voice. Every comparison ran twice with the drafts swapped. Four posts and two orders gives eight judgments for each pair.
What happened
| Comparison | Judgments won | What it shows |
|---|---|---|
| Described guide against no guide | 6 to 2 | The usual approach does help a little. |
| Two examples against no guide | 7 to 1 | Examples help more. |
| Two examples against the described guide | 8 to 0 | Real writing beat a description of it every time. |
| Two examples against the tested guide | 6 to 2 | And beat the guide we had improved. |
| Two examples against examples plus the guide | 6 to 2 | Adding the guide to the examples made drafts worse. |
| Tested guide against described guide | 2 to 6 | Our improved guide lost on posts it had not seen. |
Across the four comparisons it was in, the two-example approach won 27 of 32 judgments.
Five things we learned
1. Show, do not describe
A description turns habits into rules, and the AI then applies every rule in every paragraph. Our described guide said the writer likes short punchy lines. The drafts came back as one-line paragraphs and slogans the writer never wrote. Taylor found the same thing and called it a caricature. Two real pieces carry the proportions: how often the short lines appear, and what sits between them.
2. A guide can pass its own test and still fail
Our first rewrite of the guide lost to the original on the practice posts, 3 judgments to 5, so we threw it out. The second rewrite won 8 to 0 on the practice posts. Then it lost 2 to 6 on the four exam posts. It had learned those practice posts, not the writer. This is the reason to keep at least one piece back that nothing has been tuned on. Without the exam we would have shipped the worse guide and been sure it was better.
3. You probably have more than one voice
The judge's notes showed the writer's posts were not one style. Some use contractions all through and some use none. Some open with a personal story and some with a flat statement. A guide built from all of them averaged them into a voice the writer never uses. Recent pieces of one kind worked better than a larger mixed pile. Taylor reports the same: his 30 newest articles beat his full archive of 185.
4. Example sentences in a guide get copied
The tested guide included a sample closing line to show the tone. That exact line then turned up word for word in two different drafts. If you write rules, make up the examples, and tell the AI not to reuse them.
5. AI judges agree with themselves more than with each other
We re-ran 16 of the judgments with two smaller models. Each judge was steady: Opus gave the same winner both ways round in 23 of 24 pairs, Sonnet in 8 of 8, Haiku in 7 of 8. But on the question of described guide against tested guide, Opus preferred one and both smaller models preferred the other. A steady judge is not the same as a right one. You wrote the original, so on a close call your own ear outranks the AI's.
The easy way to do this yourself
This takes about twenty minutes and needs no software beyond a chat assistant.
- Pick three to five things you wrote in the last year or two, all the same kind: all client emails, or all reports, or all posts. Remove names and anything confidential.
- Check they sound like one person. If one is chatty and one is formal, choose the voice you want and drop the rest.
- Keep one back. That is your test piece. Paste two of the others as your examples.
- Write from full notes. Give the AI the point, the facts, the story and any quote. If your notes are a single line, the AI fills the gap with plausible content in your voice, which is the worst of both. Taylor's rule of thumb is that what you put in should be longer than what comes out.
- Run the check. Ask the AI to turn your held-back piece into plain notes, write a draft from the notes, and list how the draft differs from what you really wrote.
- Add a correction only if it earns its place. Write it as "do this, not that". Run the check again. If the draft is not closer, take the correction out.
- Feed back your edits. When you edit a draft, paste both versions and ask what you changed. Changes you make every time become corrections. Changes you made once do not.
If you want a second opinion that is not the AI marking its own work, open a new chat and paste this:
Below is something I wrote, then two drafts written by AI from the same notes. Which draft is closer to my writing in voice and style? Look at sentence and paragraph length, word choice, how formal it is, how it opens and ends, and formatting. Ignore which one is "better" and ignore small differences in content. Pick one. Then list three specific ways the other draft differs from mine, quoting a few words from each. MY WRITING: [paste] DRAFT 1: [paste] DRAFT 2: [paste]
Then do it again in another new chat with the two drafts swapped. If the answer flips, treat it as a tie.
What our test does not show
- It is small. One writer, four exam posts, eight judgments per comparison, each draft written once. A different writer could get a different order.
- The examples had an advantage. The two example posts came from the same months as the exam posts. The guides were built from posts going back two years. Some of the gap is recency, which is itself a finding, but we did not separate the two.
- We did not reproduce Taylor's method. He ran automatic rewriting over many rounds on 30 or more posts and reports large gains. We ran two rounds by hand on four. Our result says a short, hand-tuned guide overfits. It does not say his approach fails.
- AI judged AI. No person scored these drafts blind. The judges also disagreed with each other on one comparison.
- One writing model. Taylor found that a guide tuned for one AI model can do badly on another. We only wrote with one.
What the method cannot do at all is supply the ideas. It takes what you give it and says it the way you would. If the notes are empty, a perfect voice only makes the emptiness harder to spot.
Take the skill with you
We turned the steps above into one file. It builds your voice from two examples, tests itself on a piece it has not seen, writes from your notes without adding facts, and learns from your edits. It starts with no rules and adds one only after a test.
Download the skill (free)It is one plain text file. Open it, copy everything, and paste it as your first message in ChatGPT, Claude, Copilot or Gemini. Then follow what it asks. In Claude you can also add it as a skill. No sign-up is needed for this one. We checked it against our ten skill checks and walked it through one full conversation. Nobody outside our team has used it yet.More guides
Found a mistake? Use the feedback box at the bottom right. Who writes this site.