Claude AI for writing: how to set it up, prompt it, and keep the AI tells out.

1. Is Claude good for writing? Where it is strong and where it slips.

For most of the work around an article, yes. As a replacement for the person whose name goes on it, no. Claude's current Opus, Sonnet and Fable models each read up to 1 million tokens at once, which Claude's developer docs put at roughly 555,000 words, so a research pack, a style guide and five past articles fit in one conversation with room to spare. Its two newest releases were sold partly on writing: the Opus 5.5 launch post says early testers found its writing clearer and easier to follow, and the Sonnet 5.5 launch post says it writes more clearly than the previous generation. If you are still choosing an assistant rather than learning one, our Claude vs ChatGPT writing comparison runs the same brief through both.

Where Claude is strong:

• Turning a pile into a structure. Notes, interview transcripts and research go in; an outline with a clear argument comes out. This is the step most people skip and the one Claude does fastest.
• Holding long material at once. It can draft section six without losing what section two promised, as long as both live in the same conversation or Project.
• Following written rules. Rules you write down and keep in a Project get applied across every chat in it. Rules you type once in a chat are forgotten in the next one.
• Editing and rewriting. Tightening a draft, restructuring an argument, or rewriting for an audience and a reading level you can confirm with a free reading level checker.

Where it slips:

• No search data. It cannot tell you what people search for or how often. Any volume it quotes is a guess.
• A knowledge cutoff. Its current models know the world reliably to June 2026. Anything after that is missing or invented.
• Confident errors. A wrong figure reads exactly like a right one. Section eight is the fix.
• Default patterns. Left alone, its prose falls into habits readers now recognize as machine-written. Section six is the fix.
• No first-hand experience. It has never used your product or met your customers. That material has to come from you.

None of this is a secret. What separates a Claude draft you can publish from one that reads as obvious AI is how much of the second list you deal with before and after the model writes. The rest of this guide is that work, in order.

2. Which Claude model you are writing with.

As of October 3, 2026, Claude's developer docs list four current models: Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 4.5. In the Claude app, the plan you are on mostly makes the choice for you.

• Free plan: Claude Sonnet 5.5. Released September 28, 2026, it replaced Sonnet 5 as the default model on the Free and Pro plans. In the Claude apps it runs at medium effort by default, which keeps replies fast. Opus is not on the Free plan.
• Pro and above: Claude Opus 5.5. In the model menu on Pro, Max, Team and Enterprise since September 22. It is the model Claude's docs tell you to start with when unsure, and its launch post says it communicates more naturally and puts the most important information up front. Use it for long articles and the final edit.
• Claude Fable 5.1. Paid plans only, and on Pro it runs on usage credits. Worth it for long research turned into one finished document, rarely for an everyday blog post.
• Claude Haiku 4.5. The fastest and cheapest model, mostly used through the API, with knowledge that stops in February 2025. Fine for titles and summaries in bulk. Haiku 5.5 is announced but not released.

The full comparison, with API prices, speed labels, knowledge cutoffs and what one article costs on each model, is in our guide to the best Claude model for writing. If you are deciding between the two models most writers actually toggle, the Opus vs Sonnet head-to-head runs one brief through both.

One habit matters more than the pick: do not switch models halfway through a draft. Two models finishing alternate sections of one article will not sound like one writer. If you switch, switch at a stage boundary (outline, draft, edit), and keep everything else in this guide the same, because the Project and the style carry across models.

3. Set up a Claude Project as your writing workspace.

A Project is the biggest single upgrade over a blank chat. Claude's Help Center describes it as a workspace where you upload documents, give context and set project instructions that tailor every response inside it. Projects are on every plan, including Free, where you can create up to five. One Project per publication (or per client) is the right granularity.

What goes in, for a blog or a content program:

Two details from the Help Center change how you build it. On paid plans, when project knowledge approaches the context limit, Claude switches to a retrieval mode that the Help Center says expands capacity by up to 10 times. That means it searches the files instead of reading all of them, so the rules belong in the short project instructions, not on page 30 of an uploaded document. And on Team and Enterprise plans a Project can be shared with view or edit permissions, which is how a team keeps one rule set instead of five drifting copies.

Write the instructions with reasons. Claude's prompting guide notes that explaining why a rule exists helps the model apply it to cases you did not list. “No exclamation marks: our readers are finance teams and read them as hype” works better than the same rule in capital letters. The same Project, extended with keyword lists and SEO briefs, is the first setup in our Claude for SEO guide.

4. Teach Claude your voice with a custom style.

Styles change how Claude writes in any chat, inside a Project or not. Claude ships presets (Normal, Concise, Explanatory and Formal) and lets you create a custom style from the style menu in the chat box. You can build one from a writing sample you paste or upload, or from a description of how you want it to sound; Claude turns the sample into a written style description you can edit.

Use your best work, not your most recent. Give it three to five pieces that sound like you on a good day: Claude's prompting guide calls examples one of the most reliable ways to steer output format, tone and structure, and suggests three to five for best results. Vary them. A how-to, an opinion piece and a product update teach a voice; five near-identical posts teach a template, and every draft comes back in the same shape.

Then edit the description Claude writes, because it will generalize (“short sentences, conversational”). Add the specifics only you know: the words you never use, how you open a piece, how you end one, whether you write as “I” or “we”, how much you hedge. If voice-matching is the whole job, dedicated tools exist too, and our roundup of AI copywriting tools covers them.

Keep the jobs apart. The style controls voice everywhere; the Project instructions control the job (facts, links, length, banned patterns). Memory is a third layer: Claude's release notes say it has been on by default for Free, Pro and Max since late August 2026, and it picks up preferences across chats. It is useful, but it is not a style guide. Anything that has to apply every time belongs in writing, in the style or the Project, where you can read it and change it.

5. Claude writing prompts that work, one stage at a time.

The prompt that produces generic content is “write a blog post about X”. It gives Claude a topic and nothing else, so it returns the average of everything it has read on that topic. Split the job into stages, give each stage one task, and review between them. The process is the same in any assistant (we wrote up the brief-first writing process in full), so the prompts below are the Claude versions, written to lean on the Project from section three.

Three habits from Claude's prompting guide make every one of them work better:

• Long material first. Put long documents near the top of the prompt, above the instructions and the question. The guide says this improves performance across all models.
• Say what to do, not what to avoid. The guide's own example: instead of “do not use markdown”, ask for a response composed of smoothly flowing prose paragraphs.
• Write the prompt the way you want the output. The formatting of your prompt shapes the reply. A prompt full of bullet points gets bullet points back; removing markdown from your prompt reduces it in the output.

The [SOURCE?] marker in prompt 03 is the most useful line in the set. Without it, Claude fills a gap in the brief with something plausible. With it, the gap comes back visible and you decide whether to source it or cut it. Headlines are the one place to ask for options rather than one answer: ten title candidates, then pick and tighten, or compare against a headline generator.

6. How to keep AI tells out of Claude writing.

Every model has defaults, and readers have learned them. A draft can be accurate and still read as machine-written because a handful of patterns repeat across AI output, clustered in openings, transitions and endings. Claude's newest models write more naturally than older ones, but the patterns still appear, especially in long drafts and on topics where the brief is thin. The six that matter most:

1. Em dashes as the all-purpose connector. Two or three per paragraph, joining clauses that wanted a period. Readers now treat a page full of them as a giveaway.
2. Stock openers. "In today's fast-paced world", "When it comes to", "Let's dive in". They delay the answer and every reader has seen them.
3. The "not just X, but Y" construction. Once is a sentence. Three times in an article is a fingerprint.
4. Everything in threes. Three adjectives, three benefits, three examples, in paragraph after paragraph, whether or not there are three things to say.
5. Vacuous transitions. "Moreover", "Furthermore", "It is worth noting that". They signal a connection without making one.
6. Summary closers. "In conclusion", "Ultimately", "By following these steps, you can". The reader just read the article; they do not need it again.

Removing them takes three moves. First, put the list in the project instructions, phrased as what to do instead: “join clauses with a period, comma or colon” rather than a ban on a punctuation mark, “end on the next step the reader takes” rather than a ban on conclusions. Second, run the tell sweep (prompt 05) as its own pass after the edit; a model checking for one thing catches more than a model asked to write and police itself at once. Third, read the piece aloud. Rhythm problems that pass on screen are obvious out loud.

What does not work is asking Claude to “sound human” or “avoid sounding like AI”. A general instruction like that mostly swaps one set of defaults for another; naming the specific patterns is what changes the output. The test is not an AI detector score either. It is whether a reader in your niche would notice. If the post is headed for WordPress, our walkthrough on writing WordPress posts with AI without publishing slop covers the publish side of the same problem.

7. Claude for blog content that has to rank.

Writing for a blog adds a job Claude cannot do: choosing what to write. There is no keyword database inside the model. Ask it for search volume and it will produce a plausible number, which is a guess formatted as a measurement. The query, and how many people type it, has to come from real data or your own Search Console, as our guide to blog SEO explains.

Google is explicit that the tool is not the problem. Google Search Central says it focuses on the quality of content rather than how it was produced, and warns that using generative AI to publish many pages without adding value for users may violate its scaled content abuse policy. The ranking data agrees. An Ahrefs study published on July 27, 2026, covering the top ten results across 100,000 searches, found that pages under 50% AI content held 82.2% of top-three rankings, and that the average AI share rose only slightly from position one (27.1%) to position ten (30.9%). Fully AI-written pages do rank (5.3% of top-ranking pages scored 100% AI), but they are the exception. There is more of that data in our AI SEO statistics.

So the human share is the part that ranks: the query from real data, your own examples and numbers, a structure built around what the searcher needs, and the edit. Claude speeds up everything between those. Titles and meta descriptions across many posts are a good Claude job too, as long as you check length against what Google actually shows, which a meta description generator or a free SEO title generator does for you.

8. Check every claim before you publish.

Claude's developer docs are plain about it: models can generate text that is factually incorrect, and the recommended defense is to have the model cite its sources and withdraw claims it cannot support. A fluent sentence and a checked sentence look identical on the page. The knowledge cutoff makes it worse: Opus 5.5 and Sonnet 5.5 know the world reliably to June 2026, Haiku 4.5 only to February 2025, and anything newer is either missing or guessed.

The check is one prompt and one human pass. After the final edit, run the claim-list prompt (06). Then open every source yourself and cut whatever has none. Web search in the Claude app attaches citations to its answers, which makes the check faster, but a cited sentence still has to say what the page it cites says. Claims about your own product are the most common failure, which is why the facts sheet in the Project matters. This is the job an AI fact checker exists to do at scale.

A worked example: one blog post, Project to publish.

Take a fictional invoicing app, Ledgerline, writing one 1,600-word post for small agencies on how to chase a late invoice. The company, its data and its time log are invented. The setup, the prompts and the checks are the ones from this guide, run on Sonnet 5.5 on the Pro plan, with Opus 5.5 for the final pass.

Total human time: about 2 hours 20 minutes, of which the 30 minutes of Project and style setup is paid once and reused on every later post. Look at where the rest went. Drafting, the part people expect AI to remove, took 35 minutes. The brief, the tell sweep and the claim check took the other 75, and no model choice changes that.

At one post a week, that is a good trade and Claude earns its place. At one post a day it is a full-time job, and the parts that grow are exactly the ones Claude does not do: choosing the query from real data, checking every claim, and getting the finished post into the CMS. If that is where you are, the comparison worth reading is content writing priced as software against a writer or an agency.

Common mistakes when writing with Claude.

1. Starting every article in a blank chat. It feels faster than setting anything up, so the rules get retyped from memory each time and drift. Twenty minutes on a Project once beats five minutes of re-explaining every time.
2. Asking for the whole article in one prompt. One prompt is one review point, so problems surface only at the end, all at once. Brief check, outline, then one section at a time gives you a place to catch each one while it is small.
3. Building the style from the wrong samples. The most recent posts are the easiest to grab and often not the best. Pick three to five pieces you are proud of, in different formats, then edit the description Claude writes.
4. Trusting numbers Claude remembers. A specific figure feels researched. Unless it came from the brief or a source you opened, it is a guess; the [SOURCE?] marker and the claim list catch it before a reader does.
5. Telling it to sound human. The instruction is easy to write and changes little. Name the patterns you do not want and say what to write instead, then sweep for them as a separate pass.
6. Scaling with more prompts, not more checks. When the goal moves from one post a week to one a day, the tempting move is a longer prompt. The bottleneck is checking time per post, which is the whole argument for autoblogging with a quality gate rather than without one.