The short version: Claude can write blog posts that sound like you, but only if you feed it your actual sentences first, not a description of your style. Give it 400 to 600 words of your own writing, a short brief, and one blunt instruction to flag anything that reads "too smooth," and you'll get a draft worth editing instead of deleting.
Why most Claude blog drafts sound like nobody
I've read enough AI-generated blog posts now to spot one in four words. They all have the same rhythm: tidy opening, three balanced points, a soft summary at the end that says nothing. Claude is good at writing. It is not automatically good at writing like you, because by default it writes like the average of everything it's been trained on. That average is smooth, polite, and forgettable.
The fix isn't a magic prompt. It's giving Claude a sample of your real voice and telling it exactly what to copy and what to avoid. Most people skip this step because it takes five extra minutes. Those five minutes are the whole job.
The uncomfortable bit nobody tells you
Here's what people selling "AI voice prompts" don't say out loud: if your own writing is bland, Claude will faithfully reproduce bland writing in your name. A voice prompt doesn't create a voice. It copies one. If you've never had an opinion in print, don't expect Claude to invent one for you. The prompt work matters, but it's the second step, not the first. The first step is having something to say.
The setup: what to give Claude before you ask for a draft
Before you write a single prompt, collect this:
- Three to five paragraphs you've written, pasted raw, including a few of your typical quirks (short sentences, a strong opinion, a specific number)
- A one-line description of who reads the post and why they'd stop scrolling for it
- A list of five words or phrases you never use (mine includes "" and "handle")
- One real example or fact you want in the post, so Claude has something concrete to build around instead of guessing
This is the same groundwork I'd tell you to do before any AI writing session, whether that's blog drafts, tweets, or outreach. If you haven't set up a voice profile before, the approach in Claude Prompts for Editing Your Own Writing is worth doing first, because editing your own sentences teaches you what your voice is before you try to replicate it.
The core prompt
Here's the prompt structure I'd use, written so you can copy it straight in:
"Here are three paragraphs I wrote myself: [paste them]. Study the sentence length, the punctuation, where I use humour, and where I state opinions flatly instead of hedging. Now write a blog post on [topic] for [audience], in this exact voice. Keep sentences short where I would. Do not use these words: [your banned list]. Include this specific example: [your fact or scenario]. After the draft, list three sentences that sound more like generic AI writing than like me, so I can rewrite them."
That last instruction matters more than people think. Asking Claude to flag its own weak spots gets you a built-in editing checklist, and it's far more useful than asking it to "sound natural," which is too vague for any model to act on.
A worked example
Say you run a small accountancy practice and you want a blog post on why clients should switch from annual to quarterly bookkeeping check-ins. You paste in three paragraphs from an email you sent a client last month, something blunt like "Waiting until March to look at your numbers is how you end up with a surprise tax bill in April." You tell Claude the audience is small business owners who hate admin. You ban the words "streamline" and "optimise." You give it a real detail: quarterly check-ins typically catch VAT threshold issues three months earlier than annual reviews do.
Claude's first draft will likely still open with something like "In today's fast-paced business world." That's fine. You delete that line, keep the paragraph about the VAT threshold because it's specific and true, and ask Claude for a second pass: "Rewrite the opening without any version of 'in today's world.' Start instead with a short, flat statement of fact." Two or three rounds like this, each one tightening rather than regenerating from scratch, gets you a post that reads like you wrote it on a Tuesday afternoon with opinions, not one assembled by committee.
Keep a running style brief
Once you've done this a few times, save the voice description Claude gives back to you (ask it directly: "describe my writing voice in 100 words based on what you've seen") and keep that as a standing document. Paste it into every new blog post prompt instead of re-explaining yourself each time. This is the same logic behind planning content in batches rather than from scratch each week, which is covered in Q1 Content Planning Prompts for ChatGPT, even though that post is built around ChatGPT rather than Claude. The planning discipline carries across tools.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
If you want a more structured way to build prompts generally, rather than just for blog posts, the RICE Framework for AI Prompts gives you a repeatable shape (Role, Input, Constraints, Example) that works well laid over the voice-matching prompt above.
What to do with the draft once you have it
Don't publish the first output. Read it aloud. If you stumble on a sentence, that's Claude's sentence, not yours, and it needs rewriting in your own words even if the meaning is right. I'd say expect to change roughly 20 to 30 percent of the words in any Claude draft before it's publishable under your name. That's not a failure of the prompt. That's the normal ratio for AI-assisted writing that still sounds human.
It's also worth checking the draft for claims you can't back up. Claude will happily invent a statistic that sounds plausible. If you didn't give it the fact, don't trust the fact it gives back.
Different models, different jobs
Claude tends to be better than most models at holding a consistent tone across a long piece, which is why it suits full blog posts rather than short punchy copy. If you're writing social posts instead, the voice problem is the same but the fix looks different, and it's worth reading Grok Prompts for Writing Tweets That Don't Sound Like AI for the shorter-form version of this same idea. And if a blog post needs a quick first pass rather than a full voice-matched draft, a lighter model such as the one covered in How to Use Claude Haiku 5.5 can handle outlines or summaries faster, leaving the full Claude model for the voice-heavy final draft.
A simple five-step process you can reuse
- Paste three to five paragraphs of your own real writing into the chat
- Ask Claude to describe your voice back to you in plain terms
- Give the topic, audience, banned words, and one concrete fact or example
- Ask for the draft plus a list of sentences that sound too generic
- Rewrite those flagged sentences yourself, then read the whole thing aloud before publishing
Do this consistently and the gap between "AI wrote this" and "I wrote this with help" closes fast. Skip the first two steps and you'll be editing a stranger's blog post every single time.
Frequently asked questions
Can Claude learn my writing style permanently?
Not across separate chats unless you're using a feature that stores memory or you paste your style brief in each time. Treat your saved voice description as a reusable prompt ingredient, not something Claude remembers on its own.
How long should my writing sample be for Claude to match my voice?
Three to five paragraphs, roughly 400 to 600 words, is usually enough for Claude to pick up sentence rhythm, punctuation habits, and tone. Shorter samples tend to produce weaker matches.
Is it obvious when a blog post was written with Claude?
It's obvious when the draft is published unedited. Smooth transitions, hedged opinions, and generic openers are the usual giveaways. A rewrite pass that removes those and adds real specifics closes most of the gap.
Should I tell Claude to avoid sounding like AI?
That instruction is too vague to work well. It's more effective to ban specific words, insist on short sentences in places, and ask Claude to flag its own generic-sounding lines so you can fix them directly.
Run a business, a brand or a blog? I feature tools, products and expert guest posts for an audience of 15,000. See how it works
