Claude for Writing: Three Months of Daily Use
Most writing advice about AI falls into two camps: it's either "this will replace writers" or "this will ruin your voice." Three months of daily use landed me somewhere more useful: it's a genuinely good editor and a genuinely bad author.
Here's what that means in practice.
The Split: Editor vs Author
The most important distinction I found is between using the model to improve writing that exists and using it to produce writing from nothing.
As an editor — tighten this, say it more clearly, find the better word — it's excellent. It has a good ear for rhythm, spots redundancies, and catches the word that's almost right but not quite.
As an author — write this article for me — it's mediocre in a specific way: the output is competent and generic. It reads like the average of every piece of similar writing, which is exactly what it is. The voice isn't bad; it's just not yours, and it's not distinctive enough to matter.
I use it heavily for the first job and almost never for the second.
What It Reliably Improves
Clarity. This is the biggest one. When I write a sentence that makes sense to me but would confuse a reader, it's very good at flagging "this is unclear here" and offering a cleaner version. The value isn't the rewrite — it's the flag. It found the confusion before my readers would have.
Concision. Cutting a draft by 20–30% without losing meaning is something it does well, mostly by removing throat-clearing and redundant sentences I'd become blind to.
Structure. Asking "what order should these sections go in" or "what am I missing" produces genuinely useful suggestions more often than not.
Tone matching. Given a style to imitate, it's good at staying close to it. This matters for consistency across long documents or series.
Where It Hurts
It flattens distinctive voice. Left to rewrite freely, it smooths the edges — the idiosyncrasies that make writing sound like a specific person. Short, punchy sentences get expanded; unusual phrasing gets normalized. If you let it, it will iron the personality out of your writing.
It adds the same words everyone else's AI adds. Certain phrases recur across AI-generated text — "it's worth noting," "in today's fast-paced world," "delve," "underscores" — and they're a fingerprint. Over-use them and your writing signals "generated" instantly.
It's overconfident about facts. For writing that includes claims, it will state things with certainty that need verifying. The prose is confident even when the fact is shaky, which is the most dangerous combination.
The Workflow That Keeps Your Voice
After some trial and error, here's the division of labor that works for me:
- I write the first draft entirely myself. This is non-negotiable. The
first draft is where the voice and the actual thinking live; if the model writes it, both are gone and no amount of editing gets them back.
- I ask for specific edits, not rewrites. "Tighten this paragraph" or
"this sentence is unclear — suggest three alternatives" keeps my structure and choices intact. "Rewrite this for me" hands over the steering wheel.
- I read every change and reject freely. I accept maybe half of what it
suggests. The ones I reject are usually where it's sanding off my voice.
- I keep a list of my own crutch phrases. This is counterintuitive but
powerful: a list of words and rhythms that are mine helps me notice when an edit has pulled the writing away from them.
- I never let it write the conclusion. Conclusions are where judgment
lives. The model can summarize, but a real conclusion requires an opinion it doesn't have.
What Didn't Work
Drafting from scratch. Every piece I let it draft required so much reworking to sound like me that I saved no time — and the result was blander than if I'd written a rough draft myself.
"Make it sound more professional." This instruction reliably produced the worst output — stiffer, more generic, full of the tell-tale phrases. Better: "make it clearer and more direct."
Trusting it on facts. A couple of early pieces contained confident, specific-sounding errors I only caught later. Now every factual claim gets checked, full stop.
Using it for voice-sensitive work. For anything where the voice is the piece — personal essays, opinion, humor — the flattening was too costly. I use it there only for proofreading, never for editing.
Verdict
Use it as a relentless, patient editor and you'll write faster and more clearly without losing yourself. Use it as an author and you'll get competent generic text that could be anyone's.
The single rule that keeps it working: write the first draft yourself, then let it edit. Everything else follows from that.
FAQ
Will AI make my writing sound generic? Only if you let it draft or rewrite freely. Used as an editor on your own first draft, it improves clarity without flattening voice — provided you reject the edits that sand off your phrasing.
How do I avoid AI-sounding prose? Write your own first draft, ask for specific edits rather than rewrites, and learn the tell-tale phrases so you can remove them. Voice lives in the first draft, not the polish.
Is it good for editing long documents? Yes, and this is where it shines — maintaining consistency of tone and structure across something long is genuinely hard for a human and the model does it well.
Can it replace a human editor? For clarity, concision, and structure, often yes. For judgment, taste, and knowing what you're really trying to say, no — that part is still yours.