Most AI-generated text gives itself away in about four words. You've probably felt it before, even without being able to name why.
The sentences are too balanced. Every paragraph runs the same length. "Delve" shows up somewhere it has no business being. Nothing sounds like a specific person said it.
Most people try to fix this by prompting harder. "Write like a human." "Don't sound like AI." That doesn't work, because the problem isn't the prompt. Claude has never actually been told what your writing sounds like, so it defaults to the safest, most neutral version of English it was trained on.
The fix isn't a better prompt. It's a file, one you build once, that Claude reads before it writes anything for you again.
Why AI Writing Sounds the Way It Does
Every model, Claude included, was trained to be helpful, clear, and inoffensive. That training produces a specific, recognizable shape.
Predictable structure. Intro, three points, conclusion, regardless of what the idea actually needed. Real writing follows the thought, some paragraphs run one sentence, others run seven.
Safe vocabulary. Broadly understood, unlikely to offend, which is how you end up with "leverage" and "it's important to note" in places a real person never would.
No voice. AI writing is neutral by design, no pet phrases, no strong opinions, nothing that sounds like one specific person. That neutrality is what makes it feel hollow.
None of this fixes itself with a better prompt. You have to override it in writing, once.
Step 1: Interview Yourself
Open a new Claude chat and paste this:
I want to create a writing style guide that captures exactly how
I write, so I can give it to AI and have it produce content that
sounds like me. Interview me with specific questions about:
1. My sentence length preferences and rhythm
2. Words and phrases I commonly use
3. Words and phrases I never use or hate
4. How I open pieces, do I use questions, bold statements, stories?
5. How I close pieces
6. What tone I aim for (direct, warm, funny, formal, conversational?)
7. What I think makes writing good vs. bad
8. Examples of writing I admire
Ask me one section at a time. After I answer everything,
create a clean markdown file I can download and use as a
permanent reference.
Answer specifically. "I don't like formal language" gives Claude no actual rule to check against. "I never use the word 'leverage' and I hate business-speak like 'circle back'" does.
Step 2: Choosing Your Banned Words
Start with the words that are already common AI tells:
## Words and phrases to NEVER use
- delve / delving
- tapestry
- multifaceted
- it's worth noting / it's important to note
- in conclusion / to summarize / to wrap up
- leverage (as a verb)
- utilize (use "use" instead)
- robust
- comprehensive (when used as filler)
- pivotal
- crucial (when overused)
- seamlessly
- foster
- harness (as a metaphor)
- realm
- underscores (as a verb)
- furthermore
- moreover
- in today's [fast-paced / digital / ever-changing] world
- elevate (in marketing contexts)
- game-changer
- paradigm shift
- deep dive
- at the end of the dayThat's the starting list. What you add after it is what actually matters. Any word or phrase that makes you cringe in something AI wrote, add it here. Notice a construction you hate, something like "It's not X, it's Y," add the pattern itself, not just individual words.
Step 3: Write Your Positive Rules
Banning words only covers what to avoid. Here's a template for what to actually do, customize every line so it describes you specifically:
## Writing rules
SENTENCE STYLE
- Vary sentence length dramatically. Some sentences are one word.
Others run long when the idea needs space.
- Never write three sentences of the same length in a row.
- Fragments are fine when they land the right way.
STRUCTURE
- One idea per paragraph.
- Paragraphs can be one sentence long. That's allowed.
- Don't use headers unless the piece is genuinely long and complex.
- Never write a summary at the end. If it needed summarizing,
the writing wasn't clear enough.
OPENINGS
- Never start with a question.
- Never start with "In today's..." or "Have you ever..."
- Start in the middle of the thought. Assume the reader is
already there with you.
TONE
- Direct. Say what you mean.
- Conversational but not casual. Like explaining something to
a smart friend, not a stranger at a conference.
- No hedging. Don't say "it might be worth considering."
Say "do this."
WHAT GOOD WRITING SOUNDS LIKE
- Read it out loud. If you wouldn't say it that way, rewrite it.
- If a sentence could have been written by anyone, rewrite it.
- Specific details are always better than general claims.If this file could have been written by anyone, it isn't finished. It should sound like you wrote it about yourself.
Step 4: Save It and Make It Automatic
Save the file as writing-style.md. Upload it to every Claude project where you produce content, and add this line to your project instructions:
Before writing anything, read my writing style guide and apply
every rule in it. This is not optional. Every output must pass
these standards before you show it to me.
From that point on, every piece of content Claude produces in that project runs through your rules automatically.
Step 5: Set Up a Folder System, Not Just a File
A single file gets lost easily. A structure doesn't. Keep three dedicated folders wherever you work with Claude regularly.
An ABOUT ME folder, holding your voice file and anything describing who you are and how you work. A TEMPLATES folder, holding examples of your actual best past writing, real pieces, not descriptions of style. An OUTPUTS folder, where every finished draft gets saved.
If you're using Claude Cowork specifically, open the app, go to your master folder, and set this structure up on your computer directly.
Then lock the behavior in through Settings → Cowork → Global Instructions:
Read every file in ABOUT ME before every task.
Audit drafts against writing-style.md before showing them to me.
Save deliverables in OUTPUTS, never delete without approval.Step 6: Write With One Prompt, Every Time
Once the structure is in place, the actual writing prompt gets simple:
I want to [TASK] for [SUCCESS CRITERIA].After the first draft comes back:
Audit it against writing-style.md.Repeat until the flagged patterns are gone. Most drafts need one or two passes.
When the Output Still Feels Off
A word keeps slipping through. Add it to your banned list permanently instead of writing around it.
The structure still feels templated. The vocabulary got cleaned up but the shape didn't. Go back to your Structure rules and make them more explicit.
Sections still feel neutral after an audit. Your file needs more specific examples, not more rules.
You're not sure it's actually working. Ask Claude to name which specific rules it applied. If it can't, the file isn't loaded correctly.
Optional: Install the Humanizer Skill as an Extra Layer
Everything above works entirely on its own. For an additional automated pass, there's a Claude skill built for this at github.com/blader/humanizer. Where your voice file adds what's right, this skill strips out structural AI patterns as a separate, final check.
In Claude Code:
/skill install github.com/blader/humanizer
In Claude Cowork, no terminal needed:
Go to github.com/blader/humanizer, read the SKILL.md file,
and install it as a skill.
Trigger it anytime with /humanize [text]. Use it as a final pass after your voice file has already shaped the draft, not as a replacement for having one.
The Honest Part
No file makes AI writing indistinguishable from something a person wrote, every time. What this does is remove the obvious tells reliably, and give Claude something real to check against instead of guessing at "sound human."
The file is what makes the result sound like you. Everything else is a layer on top of that.
Where to Start
Never built a voice file: start with the interview prompt today, fifteen minutes.
Have a rough style guide that feels thin: revisit Step 2 and Step 3, most guides fail from a short banned list and generic rules.
Want the full system: set up the folder structure, lock in Global Instructions, use the one-line prompt from Step 6.
Want an extra layer: install the Humanizer skill as a final pass after your voice file has done its work.
The file is what makes it sound like you. Nothing else here replaces that.
If this helps something you wrote sound less like a robot, send it to one person still posting text that reads like it came out of a template.
That's all I'm asking.
Want to access all premium guides?
Become a paying subscriber to get access to this post and all other premium content.