The first genuinely useful thing Claude ever did for my writing was delete a sentence. Not rewrite it, not make it “more engaging” — just cut it, because it repeated what the paragraph above already said. Strange thing to be impressed by. But every other tool I’d tried would rather add than remove, and adding is the easy half of editing.
That tells you something about where Claude fits. It isn’t a magic draft machine. Left to its own devices, it produces prose that reads like a competent committee wrote it. What it’s unusually good at is the work around the draft: reading your stuff closely, holding an entire document in view, and telling you when your structure stops making sense. Here’s how to use Claude AI for writing in a way that actually holds up after the novelty wears off.
Where Claude earns its place in a writing workflow
It can hold an entire document at once
The context window on the current models runs to roughly 200,000 tokens, which is somewhere around 500 pages of paperback. In practice, that means you can paste in a 4,000-word draft, your notes, three competitor articles and a style guide, and still ask questions about paragraph 14. Most writing tools force you to work in fragments. This one doesn’t.
I use that for structural passes. Paste the whole piece, then ask: “What am I actually arguing here? Quote the sentence where I say it.” If the model can’t find the sentence, you haven’t written it yet.
It edits without flattening your voice
Generic AI rewrites tend to sand everything into the same smooth middle. Claude is better at doing what you ask and nothing more, which matters enormously when the instruction is “cut 20% without adding anything” or “keep my sentence rhythm, fix only the awkward clause in sentence three.”
The trick is constraining it. Vague instructions get you a competent average of the internet. Specific ones get you edits you can use.
It’s a decent skeptic
My favourite prompt of the last year: “Here’s my argument. Give me the three strongest objections a smart reader would raise, then tell me which one I haven’t answered.” It’s not a substitute for a real editor, but it catches the holes you’ve stopped seeing because you wrote the thing.
Prompts that beat “write me a blog post about X”
Generic drafting prompts get generic drafting. These do more work:
- The diagnosis: “Summarise my argument in two sentences. Then tell me which paragraph doesn’t serve it and why.”
- The voice capture: “Read these three samples of my writing. Describe my sentence length, vocabulary and verbal tics in specific terms. Use that description as a style guide for the rest of this conversation.”
- The buried lede: “I’m writing for [specific audience]. What’s the one thing they need first that I’ve hidden in paragraph five?”
- The pointed critique: “Don’t rewrite anything. Just name the weakest sentence in each paragraph and say in one line why it fails.”
- The tone ladder: “Rewrite this paragraph three times — plainer, warmer, sharper — and leave the original intact so I can compare.”
Notice how few of these ask for new prose. The writing is your job. The reading, checking and reframing is where the leverage is.
Where it slips
Be honest about the failure modes, because they’re consistent. Claude hedges. It loves a neat list of three. It likes to close a section by restating what the section just said, which inflates word counts and dulls rhythm. It also mirrors your own clichés back at you — if you write “it’s worth noting,” you’ll get more of it.
The bigger risk is confident invention. Dates, quotes, statistics, who said what in which report — these can arrive sounding authoritative and be entirely wrong. Treat every number and quotation as unverified until you’ve found the source yourself. The Claude AI chatbot has a longer list of strengths and blind spots worth knowing before you build a daily habit around it.
A workflow that consistently works
Step one: make it interview you
Before any drafting, open with: “Ask me ten questions about this topic before you write a word. Focus on what I know that isn’t in the top search results.” You’ll get questions about your own experience, the numbers you have, the objections you’ve heard. Answer them badly and quickly — nobody sees it. What you end up with is material that couldn’t have been generated, because it came from you.
Step two: outline in one message, prose in another
Agree the structure first. Argue about the order of sections while it’s still cheap to change. Then draft straight through against that skeleton. Editing an outline takes minutes; unpicking 1,500 words built on a bad spine takes an afternoon.
Step three: draft and edit in separate conversations
This one matters more than it sounds. A fresh chat has no loyalty to the draft you just wrote. Start a new conversation, paste the piece in cold, and ask for the structural pass. The feedback is noticeably harsher, which is exactly what you want from an editor and exactly what you don’t get from an assistant that watched you build the thing.
How much Claude you actually need for writing
Long-document work is the resource-hungry part. Pasting 5,000 words and running several passes will eat through a free allowance quickly, and the free tier throttles you after a modest number of messages — the specifics of what the free plan gives you and where it runs out are worth reading before you plan a workflow around it.
For most people writing regularly, the $20 tier covers it comfortably. If you’re processing book-length manuscripts or running dozens of long-context passes a week, the higher tiers start to make sense. There’s a clear breakdown of what each Claude plan includes at the $0, $20, $100 and $200 levels if you want the numbers rather than the marketing.
And if you’re still deciding which assistant deserves the subscription, this honest comparison between Claude and ChatGPT from someone who pays for both is more useful than a spec sheet.
Keep the parts that only you can write
There’s a temptation, once the drafting gets fast, to let the drafting get thin. The paragraph about the client who changed their mind halfway through, the number you pulled from your own analytics, the opinion you’d defend in an argument — those are the bits that make an article worth reading, and they can’t come from a model that has never met you.
So arrive with material. Notes, figures, a half-formed gripe about how everyone in your industry does the same thing. Claude is genuinely excellent at turning that into something readable, restructuring it, cutting the fat and telling you where the logic breaks. It is not much good at inventing the material, and it will happily fake it if you let it.
The writers getting the most out of this aren’t the ones with the cleverest prompts. They’re the ones who show up with something to say, and use the tool to say it faster and with fewer loose threads. Gathering the material was always the hard part. That hasn’t changed.