How to Make Claude Write Like a Human: The Full Method

Build a voice file once, reuse it forever

You paste a draft into your editor, read two paragraphs, and delete the whole thing. Balanced sentences. Three tidy bullets. A wrap-up about where all this is headed. Claude wrote it, and every reader will notice within ten seconds.

If you want to know how to make Claude write like a human, the honest answer is unglamorous: not a prompt, a file. One document, usually Markdown, that captures who you are on the page. Your samples, your tics, your rules.

Why does a file work when prompts don't? Because by default Claude writes for an average reader, and averages read like press releases. Anthropic's prompting guidance (be direct, name your audience and format up front, skip the preamble) strips padding — but stripping padding isn't the same as adding voice. A three-sentence prompt can't carry a voice. Examples can.

Here's how to build yours, step by step:

1. Collect 3-5 real samples. Pick emails, posts, or articles you wrote yourself with zero AI help. The more casual the better. A text message and a ranty Slack message count.
2. Describe your own habits. Open a sample and ask: do I write short paragraphs? Do I use parentheticals constantly? Do I swear, joke, digress?
3. List your banned words. More on this below, but capture them now while you're in the file.
4. List your go-to phrases. Words and expressions you use constantly — "honestly," "here's the thing," starting sentences with "And."
5. Write your structural rules. Paragraph length, how often you use lists, whether you address the reader as "you."
6. Save it as a .md file and paste it at the start of any writing session, or install it as a Claude Skill so it loads automatically.

What goes in the banned-words list

Every writer's list is personal. The AI tells, though, are common enough that a starter list covers most of the damage.

Start with the openers: the ones that announce a topic instead of starting it, the ones that frame everything as a grand "world" you're about to enter, the ones that delay the point with a pivot phrase. Delete on sight.

Then the emphasis words: anything promising a breakthrough, plus "crucial," "pivotal," "seamless," "robust." Readers have learned to flinch at them.

Corporate verbs next. "Leverage," "unlock," "elevate," "harness." A person says "use." A press release says "leverage."

And the closers: formal sign-offs that announce a conclusion is arriving, futurist flourishes about where all this is headed, any line promising that everything you just read is only the beginning.

The rhythm tells are harder to list because they're structural: paragraphs of identical length, three bullets every single time, a summary that restates the article. You'll catch those in the second pass.

If you only have time for one part of this, make it the banned list. Samples teach Claude what to imitate. The banned list tells it what to stop doing, and stopping is where most of the humanizing happens.

The second pass: Claude critiquing Claude

This is the technique most guides treat as a footnote, and it makes the biggest difference. Drafting, critiquing and revising is a loop you can run explicitly instead of hoping the first draft is right.

After Claude writes a draft, don't edit it yourself. Ask Claude to attack it:

"Read back what you just wrote. Highlight every word or phrase a human would never say out loud. List them. Then rewrite the piece, replacing each one. If you're not sure whether a phrase sounds human, it doesn't."

Run this once and you'll usually get a handful of flagged phrases per few hundred words. Then paste in one of your real samples and ask "what in this draft still doesn't match this voice?" for another round. Two passes is the sweet spot. Three and the text starts to flatten out, losing the unevenness that makes writing feel alive.

That last row is underrated. Claude pads. Telling it to cut a percentage forces compression, and compression is what human writing looks like under pressure.

The humanizer skill: worth installing?

There's a well-known open-source skill called humanizer (by blader on GitHub) that targets a long list of specific AI patterns by name and runs a second pass asking "what still makes this obviously AI-generated?" before fixing it. It automates the critique loop above without you typing the prompts each time.

If you go this route, you're in the market for Claude Skills, and it pays to know what else is out there. Our take on the best Claude Code skills covers which ones earn a permanent slot versus which ones you try once and uninstall.

The honest assessment of humanizer specifically: it's good at removing tells, weaker at adding voice. It strips the machine smell but leaves neutral prose behind. Pair it with your voice file or you'll get clean, boring text instead of human text.

One caution. Skills that rewrite aggressively can overcorrect, turning your draft into something that tries too hard to sound casual. Humans vary. Some of your sentences should be formal. Don't let a skill sand that off.

Prompts that actually work (and one that doesn't)

The prompt that doesn't work: "make this sound more human." It returns the same text with contractions added.

What works instead:

• Sample-anchored rewrite: "Rewrite this in my voice. Here are two things I wrote recently. Match the rhythm, not just the words."
• Speak-it test: "If a sentence wouldn't survive being said out loud to a friend, rewrite it."
• Imperfection pass: "Vary my sentence lengths. Make some paragraphs one line. Let one thought end abruptly."
• Perspective injection: "Here's my actual opinion on this: [state it]. Weave it in. Don't neutralize it."
• Reader-specific framing: "Write for a skeptical developer who's seen 50 AI listicles this month" produces sharper prose than "write an article."

The perspective prompt is the sleeper hit. Claude hedges by default. Hand it a genuine opinion and it finally has something to argue for instead of around. Every prompt on this list shares that trait: it teaches Claude something instead of asking for a mood.

Does the model itself matter?

Less than the setup, but not zero. Larger Claude models tend to hold a voice across a long piece better than smaller ones, and a large context window means you can drop several long writing samples into the conversation without a second thought.

Don't agonize over plan tiers while your voice file is still empty — check Anthropic's current pricing when you need to, but voice quality comes from your setup, not from a bigger quota. The model is maybe a tenth of the answer; the other nine-tenths live in the file and the critique loop.

The honest limits

Some things no amount of prompting fixes, and you should know them before you invest an afternoon:

• Claude can't invent your experiences. If the piece needs a story that actually happened to you, Claude will fabricate a plausible-sounding one. Supply the anecdote yourself, always.
• It converges back to default. Long sessions drift. Re-paste the voice file every few thousand words or start a fresh chat with it loaded.
• Detection tools still flag polished prose. Uniform sentence rhythm survives a lot of humanizing. Deliberately breaking rhythm — one-line paragraphs, a fragment now and then — does more than any banned-words list.
• It won't sound like you on the first pass. Expect three or four rounds before the voice file is dialed in. That's normal.

How to make Claude write like a human: the full loop

Expect the first version of your voice file to be short and partly wrong. You'll describe the habits you know about ("I write short paragraphs") and miss the ones you don't — the digressions, the mid-sentence asides, the way you actually open an email. Only your own drafts can teach you those.

So treat it as a loop:

1. Draft with the voice file loaded and the audience named.
2. Hunt for AI patterns with the critique prompt.
3. Add every phrase that slipped through to the banned list.
4. Compare against a real sample and fix what still doesn't match.
5. Watch for overcorrection. If the prose starts reading like a parody of you, leave a few formal sentences alone.

The file is never finished. Every draft teaches it one more thing about you.

That's the whole answer to how to make Claude write like a human: a file that keeps getting more honest, a list that keeps getting longer, and a critique loop you run every single time. A reader who knows you should finish a paragraph and think "yeah, that's them." A stranger should just finish it.

FAQ

How do you humanize text using Claude?

Paste a voice file with 3-5 real writing samples at the start of your session, then ask Claude to highlight every phrase a human would never say and rewrite the piece. Two critique passes beat any single "make it human" prompt.

Does Claude write like a human by default?

Partially. Its default register is an average of all readers, which reads as generic. Voice samples and a banned-words list close the gap.

How do I set up Claude to write like me?

Create a voice file with your real writing samples, habits, banned words, and go-to phrases. Save it as a .md file, install it as a Claude Skill, and re-paste it every few thousand words so it doesn't drift back to default.

What is the best humanizer skill for Claude?

The open-source humanizer skill by blader targets a long list of AI patterns by name and runs a second critique pass. It's good at removing tells but doesn't add voice, so pair it with your own voice file.

How do I keep my writing skill on every machine I use?

Install it once through loadout. The agent on each paired machine installs the same skill, so the voice file you tune on your desktop is already on your laptop.

One last thing

A voice file and a humanizer skill are part of your setup — and setups drift. Tune them on your desktop, sit down at your laptop, and none of it followed you. That's a sync problem, not a writing problem, and it's the one loadout handles: your Claude skills and MCP servers stay in sync across every paired machine from a single dashboard. If you're still choosing the rest of your stack, start with our ranking of the best MCP servers for Claude Code, or see how syncing works.