# Teach the AI to write like you, not like everyone > Every AI draft sounds the same because the model never read you. The fix is four plain files and one editing skill, built from your own old writing. The full method, with the prompts, deep enough to copy. - Source: https://athr21.com/writing/teach-the-ai-to-write-like-you/ - Published: 2026-09-01 - Track: Field notes - Reading time: 10 minutes - Language: English - Author: Aly Zeineldin, Athr21 --- Stop writing with AI. At least the way most people do it. Not because the tool is bad. Because everyone who writes with it has started writing the same piece. The same punchy opener, the same "it's not just X, it's Y" line, three bullet points, an em dash in the middle of every second sentence, and a closing paragraph that repeats what you just read. Readers learned the pattern in about a year. Now the pattern is the reason they scroll past. I write with AI every day. Proposals, reports, articles like this one, my LinkedIn posts. Almost nobody notices, and that is the point. The difference is not a better prompt. "Write in a confident, casual tone" describes ten million people. The only thing that describes you is what you already wrote. So I gave the AI that, and I gave it a filter for the habits it cannot drop on its own. This is the whole method, in the order it was built, with the prompts. It needs no code. It needs your old writing and about two hours. ## Why the average sounds like the average A language model predicts the most likely next word for the widest range of readers. Left alone, it writes the mean of the internet. That is useful for a summary and fatal for a voice, because a voice is exactly the set of choices that are not the mean. You cannot prompt your way out of that. You can only give the model something more specific than the internet to predict from: you. ## Step 1: collect everything you wrote before AI Make one folder. Mine is called `voice-samples`. Put in it every piece of writing you produced by hand, before you started using AI:
Posts
Your LinkedIn posts. Export them from Settings, or copy the ones you are proud of by hand.
Emails
Ones you wrote that got the reply you wanted. Strip anything confidential first.
Notes and talks
Old blog posts, a talk you gave, transcripts of voice notes you sent a colleague.
Both languages
If you write in Arabic and English, collect both. The voice is different in each and the AI needs to see that.
Two rules. Only things you wrote yourself, and only from before you used AI, otherwise you are teaching the model to imitate itself. Aim for twenty pieces or more. The archive is the asset. Everything below is built from it. ## Step 2: distil a voice file from the archive Give the AI the whole folder and this prompt. It produces one page that describes how you write, with evidence. Read it, cut what is wrong, and keep it as a file the AI reads before every draft. ```text Read every file in this folder. They were all written by me, by hand. Write a one-page "voice file" that describes how I write, with examples quoted from the files. Cover: 1. The base: language, dialect, which English terms I keep untranslated. 2. Signatures: five to seven things I do in almost every piece. How I open, how I close, my rhythm, how I use lists, punctuation habits. 3. Skeleton: the structure of a typical piece, line by line. 4. Words and constructions I never use. 5. Ten reference lines quoted from my own writing that sound most like me. Do not flatter me. Do not describe a tone I should have. Describe the tone I have, with evidence. ``` Mine, for Egyptian Arabic, came out as a page. Shortened, it reads like this: spoken Egyptian with English terms left in Latin script, not formal Arabic and not translated English. A hook in line one that flips what the reader expects. Operator language, the P&L and the system, married to warmth. Reframes by contrast. Short sentences, one idea each. And ten lines quoted from my own posts that the AI can hold as a tuning fork. The last section matters more than it looks. A description of a voice is a guess. Ten real lines are a standard. ## Step 3: write the banned list The voice file says what you do. This one says what the machine does and you do not. Mine is called `Writing Baseline.md`, it applies to every project I run, and it is deliberately short. Copy it as it is, then add your own.
Banned by default
Inflated transitions: moreover, furthermore, additionally, "in today's fast-paced world".
The rule-of-three reflex.
Hedging filler: "it's important to note", "plays a crucial role".
Empty intensifiers: very, truly, genuinely, "a testament to".
"Not just X, but Y." "It's not about X, it's about Y."
Summary closers: in conclusion, ultimately, "at the end of the day".
Em dashes as the default rhythm.
Promotional adjectives instead of evidence: seamless, robust, game-changing.
Title Case Headings.
Do this instead
Lead with the point. No throat-clearing.
Concrete over abstract: the number, the company, the date.
Short first drafts I can sharpen. I edit heavily.
One idea per sentence. Vary the length so it reads like speech.
Active voice, real subjects. "We shipped", not "it was shipped".
Never invent a figure. If unverified, say so.
## Step 4: install the humanizer This is the piece people ask about. It is a skill: a text file of editing instructions the AI loads whenever it finishes a draft. It is built on Wikipedia's page called Signs of AI writing, maintained by the editors whose job is cleaning AI text out of Wikipedia. It lists thirty-three patterns, each with a before and an after. On every draft it runs the same loop:
1 · Scan
the draft for the 33 patterns
2 · Rewrite
keeping every paragraph and point
3 · Ask
"what still reads as AI?" and list it
4 · Rewrite again
until the list is empty and the em dashes are gone
5 · Match
the voice file, not a generic "natural" tone
The thirty-three patterns fall into a handful of families. These are the ones that catch most drafts:
Inflated significance "marks a pivotal moment", "underscores the importance of". State the fact and drop the ceremony.
The -ing tail "..., highlighting the need for", "..., ensuring that". Full stop. New sentence, or cut it.
AI vocabulary delve, crucial, landscape, tapestry, vibrant, showcase, foster. Use the plain word.
Avoiding "is" "serves as", "stands as", "boasts". It is. It has.
Negative parallels "it's not just about X, it's Y". Say Y.
Dressing em dashes, bold on every phrase, an emoji on every bullet, curly quotes, Title Case. Commas, full stops, sentence case.
Chatbot residue "I hope this helps", "let's dive in", "great question". Delete.
Manufactured drama stacked one-line punchlines, "X is the currency of Y". Write the concrete claim underneath.
Fake candour "Honestly?", "Here's the thing", "Real talk". Just say the thing.
Where it lives depends on your tool. In Claude Code or Cowork, save it as `~/.claude/skills/humanizer/SKILL.md` and it loads when needed. In ChatGPT, paste the same text into a Project's instructions. With nothing fancy at all, keep it as a document and paste it in with each draft. One more line made the real difference. I told the AI, as a standing rule in its memory: run the humanizer on every draft, without being asked, before I see it. A step you have to remember is a step you skip. ## Step 5: the context that makes it you Voice is half. The other half is knowing what you are talking about. My [second brain](/writing/build-your-second-brain-in-7-steps/) holds a few plain text files the AI reads at the start of every session:
About me
Who I am, what I do, how to talk to me, and what never goes public.
Project notes
One per project: the rules, the decisions already taken, and why.
The brand file
Colours, tone, the banned words, the one call to action.
Voice and baseline
Steps 2 and 3 above. The last two files the AI reads before it writes.
Five text files. No subscription and no new tool. The same files power proposals, the website and email, which is why the two hours are worth it: you set the layer up once and every output after it inherits the voice. ## Step 6: measure, or the voice drifts back Sounding like you is the floor. So I keep one more file: a short list of writing principles the AI reads before drafting, and each line carries a status. Seed means untested. Proven, weak, or killed mean the numbers have spoken. Every draft ends with a small block naming which principles it applied and which one it is testing. Every few days a scheduled task pulls the results and grades the list. The list changes what the AI writes next, which is the difference between a style guide and a learning loop. A style guide is what you believed on the day you wrote it. The ledger is what actually held up. ## Honest limits This is not training a model. It is context. The AI forgets it the moment the files are not in the room, which is why the files live in one place it always reads. It still drifts. Left alone for a few pieces it slides back towards the average. The humanizer catches most of it. Your edit catches the rest, and if you stop editing, the voice goes with it. And it will never have a story you did not live. The archive gives it your rhythm. You still have to give it the thing that happened this week. ## Do it yourself If you already work inside an AI workspace, paste this and start: ```text I want the AI to write in my voice, not the average internet voice. Work through these stages in order and wait for my review between each: 1. ARCHIVE. Ask me for a folder of writing I produced by hand, before I used AI: posts, emails, notes, in every language I write in. Read all of it before anything else. 2. VOICE FILE. From the archive, write a one-page voice file: the base (language, dialect, terms I keep in English), five to seven signatures with quoted examples, the skeleton of a typical piece, the words I never use, and ten reference lines quoted from me. Describe the voice I have, not one I should have. 3. BANNED LIST. Write a short writing baseline: the AI habits I never want (inflated transitions, rule of three, hedging filler, empty intensifiers, "not just X but Y", summary closers, em dashes as rhythm, promotional adjectives, Title Case) and the plain habits I want instead. 4. HUMANIZER. Install an editing pass based on Wikipedia's "Signs of AI writing" page: scan every draft for its patterns, rewrite without losing any point, ask "what still reads as AI?", rewrite again, and match my voice file. Zero em dashes in the final. Save it as a skill or as standing instructions, and run it on every draft before I see it, without being asked. 5. CONTEXT. Keep the voice file and the baseline next to my About Me, my project notes, and my brand file, and read all of them at the start of every session. 6. LEDGER. Keep a principles file with a status per line (seed, proven, weak, killed). End every draft with the principles it applied and the one it is testing. When results come in, update the statuses. Rules: never publish anything on my behalf. Never invent a number or a quote. Keep drafts short enough for me to sharpen in ten minutes. ``` The prompt is the compressed version of what took me a few weeks to settle. It will not remove the editing, and it should not. The editing is where the voice lives. The deeper point stands without any AI. Write down how you write, what you never say, and what actually worked, and every draft after that starts from you instead of from the average. That is the same discipline [The Brain](/services/#brain) installs inside a company, for the whole team, not one writer.