The short version: Claude writes better content than most AI tools because it holds context, tone and a proper brief better than ChatGPT, but it will still write generic rubbish if you feed it a generic prompt. The fix is a Project with your actual voice and examples loaded in, a brief that tells it what to leave out, and you editing the output rather than pasting it straight to WordPress. None of that is complicated. Almost nobody does it, which is why so much AI content still reads like it came out of a machine.
Useful alongside this: How AI Tools Can Help You Write a Research Paper Faster (Without Getti.
Why Claude, specifically
I use all three of the big models for client work. ChatGPT is faster to bounce ideas around in. Gemini is handy when I need something pulled from a live Google Doc or spreadsheet. But when it comes to the actual writing, the sentence-by-sentence craft of it, Claude is the one I open first.
Here is why, in plain terms. Claude tends to hold a brief across a longer conversation without drifting back to its default corporate cadence. Ask ChatGPT to write in a punchy, first-person voice and by paragraph six it has quietly slid back into "in today's fast-paced digital landscape" territory. Claude does this less. Not never, but less. It also handles nuance in instructions better, things like "don't use the word unlock" or "keep sentences under 20 words in this section," and it keeps to them for longer.
That said, I want to be honest about something most posts on this topic skip over. Claude is not a writer. It is a very good pattern-matcher wearing a writer's coat. It cannot know your customers, your objections handled on sales calls last Tuesday, or the specific thing your competitor got wrong that you can quietly needle in a paragraph. The tool does not make you a better writer. It makes an average writer's output faster and a bad writer's output invisible, because bad writing dressed up in confident AI sentences still reads confidently. That is the uncomfortable bit nobody selling AI writing courses wants to say out loud. The skill you need is not prompt writing. It is knowing your own content well enough to spot the fifteen seconds where Claude got it wrong.
Step one: stop using the free chat window for real content
If you are pasting a topic into Claude's regular chat box and hitting send, you are using maybe a third of what the tool can do. Claude.ai has a feature called Projects (on the Pro plan, roughly £16 to £18 a month depending on when you check, or Team plans if you've got people helping you). A Project lets you upload documents that sit permanently in the background of every conversation in that Project. No re-explaining yourself every session.
What I load into a client's content Project, every time:
- A one to two page brand voice document: how they open sentences, words they use and words they ban, three examples of a piece they loved and why
- Five to ten of their best-performing existing posts or emails, pasted in full
- A one-paragraph description of their actual customer, not a demographic, an actual person with an actual objection
- A "banned words" list specific to that brand, on top of the generic AI tells like "delve" and "unlock" and "in today's world"
Do this once and every piece of content you generate in that Project starts from a much higher baseline. If you haven't put together that voice document yet, I've written a full walkthrough of how to build the brand voice document that stops AI content sounding generic, and it's the single highest-use 45 minutes you can spend on this whole process.
Step two: write the brief like you're briefing a junior writer, not a search engine
Most people's Claude prompts read like SEO briefs: "write a 1500 word blog post about email marketing for small business owners." That gets you a serviceable, forgettable article. Every competitor is getting the exact same output from the exact same prompt.
Here's what I send Claude for a client blog post, roughly, stripped down for one recent job in the beauty salon software space:
- The real question the reader is typing into Google, word for word
- Two things every other article on page one gets wrong or skips (I check the top five results first, this takes ten minutes)
- One specific number, tool name, or real example I want included, so it's not floating in the abstract
- The angle the client believes, even if it's slightly unpopular, e.g. "most salons don't need a booking app, they need to stop double-booking on a Saturday, which is a staffing problem not a software problem"
- A closing instruction: "don't summarise what you just said in the last paragraph, end on the specific next step"
Notice there's no instruction telling it to "sound human." That instruction does almost nothing on its own. What changes the output is giving it specific facts and a specific opinion to build the piece around, because generic prompts can only produce generic content, and that's true whichever model you use.
A real example: the newsletter that got rewritten three times
Last year I was working with a small consultancy client on a weekly newsletter. First draft from Claude, decent brief, brand voice document loaded, came back at 640 words and read fine. Grammatically perfect. Nothing wrong with it. And nothing anyone would remember five minutes after reading it.
Second pass, I gave Claude one instruction: "take out every sentence that could have been written about any consultancy in any industry." That single prompt cut the piece by about a third. What was left was thinner but sharper, because it forced the model to keep only the sentences tied to a specific fact or specific client story I'd fed it earlier in the conversation.
Third pass, I added back one thing myself, by hand, that Claude could never have known: a throwaway line about a client call that morning where someone had asked the exact question the newsletter answered. Thirty seconds of typing. That one sentence is the reason that newsletter got more replies than any edition in the previous two months. Claude wrote 90 percent of the piece. The 10 percent I wrote is the reason it worked. That ratio, roughly, is what I'd expect anyone doing this well to end up with, not because Claude is bad at its job but because the last 10 percent is the bit that only you have access to.
Sonnet versus Opus versus which model to pick
People overthink this. For most content writing, Claude's Sonnet model (the mid-tier one, included in the standard Pro subscription) is what I use for probably 80 percent of drafting work. It's fast, it's cheap on your usage limits, and for blog posts, emails and social captions the difference against the top-tier Opus model is marginal.
Where Opus earns its higher cost is longer, more analytical pieces, things like a 4000-word cornerstone guide where the model needs to hold a structure across many sections without losing the thread, or a piece where you're asking it to reason through a complicated argument rather than just execute a clear brief. For a straightforward "write me a LinkedIn post about this stat," you're wasting nothing but time reaching for Opus.
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.
The mistake that makes AI content sound like AI content
I read a lot of client content before they hire me, and the tell is almost always the same: every sentence is the same length, every paragraph makes exactly one point and then restates it, and there's a summary at the end that repeats the introduction using different words. None of that is Claude's fault specifically, it's what happens when a first draft goes out as a final draft.
The fix costs you maybe eight minutes per piece:
- Read it out loud, once, and cut anything you'd never say to a person's face
- Delete the last paragraph if it just repeats the introduction, replace it with one specific action
- Vary sentence length on purpose, a four-word sentence after three long ones does more for "sounding human" than any prompt instruction ever will
- Add one real detail Claude could not possibly have invented, a date, a client name (with permission), a number from your own data
This is the same reason I keep saying the model matters less than the person editing it. If you want a deeper look at exactly what's giving AI writing away at the sentence level, I've broken it down in why your AI content sounds like everyone else's (and how I fixed mine).
Turning what you already say into Claude prompts
The easiest content to make sound like you isn't content you type from scratch, it's content built from things you've already said out loud. If you record a 15 to 20 minute voice note walking through a client problem, an opinion, or how you'd solve something, and drop the transcript into Claude with an instruction to "turn this into a blog post keeping my phrasing wherever it's usable," you skip most of the generic-AI problem entirely, because the raw material was never generic to start with. I use this constantly and wrote a full process for it in the voice note content pipeline, it's the fastest way I know to get a week of content that still sounds like a real person said it.
Where "better content" and "content that sells" split apart
Here's a distinction most guides on this topic miss entirely. Claude can absolutely help you write content that ranks, reads well, and gets shared. It cannot on its own write content that converts, because conversion depends on things outside the text: your offer, your pricing, your proof, your follow-up sequence. I've seen client blogs climb to page one with beautifully "Claude-improved" writing that still does nothing for revenue, because the article answered the question and then just... stopped, no path to the next step. I wrote about this exact trap in why AI blog content ranks but doesn't sell, and it's worth reading before you spend another month polishing prose instead of fixing the offer sitting underneath it.
Training Claude on your voice, not just once
A voice document isn't a one-off task you tick off and forget. Voice drifts, your business changes, you find better phrases in a sales call six months from now that should get fed back in. I revisit the Project files I use with clients every few months, adding fresh examples and cutting old ones that no longer represent how the business talks. It takes maybe 45 minutes each time and the improvement in output is noticeable within the first few generations. If you want the full method for that initial setup, including exactly which documents to prioritise and in what order, I've laid it out step by step in the 45-minute fix that stops your AI writing like everyone else's.
A short, honest checklist before you publish
- Did you upload real examples of your own writing before generating this, or start from a blank prompt?
- Is there at least one specific fact, number or story in the piece that Claude could not have invented on its own?
- Did you read it out loud once, all the way through?
- Does the ending tell someone what to do next, rather than summarising what they just read?
- Would this paragraph make sense if you swapped your company's name for a competitor's? If yes, cut it.
Run every piece through that before it goes anywhere near a publish button. It takes ten minutes and it's the difference between content that reads like it was written by a person who knows the subject, and content that reads like it came from a very confident autocomplete.
Related reading: data make better blogger.
Frequently asked questions
Is Claude better than ChatGPT for writing blog posts?
For holding a specific tone and brief across a longer piece, most writers I work with, myself included, find Claude more consistent than ChatGPT, which tends to drift back toward generic corporate phrasing the longer a conversation runs. ChatGPT still has the edge for quick brainstorming and live research. Neither will write good content off a lazy one-line prompt.
Do I need the paid version of Claude to write good content?
You can get useful drafts on the free tier, but the Projects feature, which lets you permanently load your brand voice document and past examples, sits behind the Pro plan (roughly £16 to £18 a month). That feature alone is what separates generic-sounding output from content that sounds like your business, so for anyone writing content regularly it pays for itself within a couple of articles.
How long should a good Claude prompt for a blog post be?
Longer than most people write, and shorter than most people think. A strong brief usually runs 100 to 200 words: the exact reader question, one or two things competitors get wrong, one specific fact or number to build around, and your actual opinion on the topic. A one-line topic prompt with no specifics is the single biggest reason AI content ends up sounding the same everywhere.
Can Claude write content that sounds exactly like me?
Close, but not exactly, and that's the honest answer. It can match your sentence patterns, common phrases and structure convincingly if you feed it real examples of your writing. It cannot know the specific client story, the throwaway detail from a call this morning, or the slightly unpopular opinion you hold, unless you tell it. The last 10 percent of any piece, the part that makes it unmistakably yours, still has to come from you.
Related reading: Best Interactive Content Tools in 2026 and The Best AI Writing Tools for Bloggers in 2026 (What I Use, Not What I'm Paid to Say).
