At 8:12 on Tuesday, I was supposed to be halfway through a client’s newsletter. Instead, I was reading my opening from the night before — “Welcome to another edition, packed with helpful tips to improve your home and lifestyle” — and wondering when I had become a brochure.
The client runs a small home-care service — visits to elderly customers, cleaning, shopping, checking nobody has quietly stopped eating properly. The newsletter had to announce a new Saturday visit without sounding like an advert for loneliness, ready by 10:30.
I opened Claude first. Habit, not a testing system — its drafts come back slightly battered and recognisably alive.
Claude wrote like a person. Mostly.
I gave it the facts: Saturday visits from 9 to 1; no long contract; the same carer where possible; and families booking for a parent without the parent feeling “inspected” — the founder’s word. Also, from her voice note:
“Mostly, we are another pair of eyes, because families cannot pop round every day.”
Claude’s first paragraph came back:
“Somebody who knows your mum’s routines can notice when something has changed — the milk lasting too long, the curtains still closed, the lunch untouched.”
Better than anything I’d written at midnight. Specific. Quiet. No hard sell. This is why I say Claude sounds more human: the rhythm is already in the first draft — a short sentence where one is needed, then a longer one that wanders a bit and lands.
Then it spoiled the run:
“Together, we can create a brighter, more nurturing home where every precious moment truly matters.”
Precious moment. Truly matters.
I replaced it with: “The visit is practical first: shopping, cleaning, a cooked lunch, and somebody checking that the house feels lived in.” Less uplifting. Much better. Claude can be warm without being salesy — unsupervised, though, “warm” tips into greeting card.
The rest held together — the point stayed “another pair of eyes,” not a services list. On longer jobs (a 4,000-word guide this spring), that thread-holding matters even more.
Then I needed the annoying parts done quickly
The manager — Sadia, who runs the rota and fears neither God nor deadline — needed a heading, three subheadings, two subject lines under 40 characters, a WhatsApp version, and a caption that did not repeat the newsletter. None of it difficult. All of it fiddly. You know the sort: nothing you’d ever put in a portfolio, all of it somehow urgent. And the clock said 9:10. Eighty minutes left.
ChatGPT is not who I’d pick to write the soul of the piece. It’s who I pick when the piece needs twenty doors cut into it before lunch. I asked for headings in the house style, exactly three subject lines, no exclamation marks, and a 35-word summary. It obeyed. Instantly.
It also threw off a pile of captions — twelve? fourteen? I stopped counting — while I hunted for the logo (folder: “final logos FINAL”; it is never the final one). Most were interchangeable. Two usable. One was right after I removed “peace of mind,” a phrase clients love and I hate, because it promises something no visit can guarantee.
Its first heading, though, was this:
“Unlock a Safer, Happier Home in Just Three Simple Steps!”
The complete starter kit: unlock, a promise, a number, and the exclamation mark I had just banned. I changed it to “What the Saturday visit actually includes.” A product page might need the formula. A note to worried families did not.
On a deadline, the difference is not subtle — no, that’s not quite it either. It’s not the deadline. It’s that they’re not even doing the same job. Claude gave me the voice. ChatGPT gave me — well. Twelve-ish captions, three subject lines, one heading I had to rescue from an exclamation mark. Variations, if we’re being generous. I needed both. Just never for the same sentence.
The edit that told me more than a week of testing
Around 9:40, the client sent a paragraph she had written herself. Clunky, repetitive, and containing the best line in the newsletter:
“My mother would rather set fire to the calendar than ask me for help.”
She wanted it “tidied up” without losing her voice — a dangerous request, because tidying can mean sanding the person off the page.
I tried Claude first: “Edit this, but keep my wording where it works.” It removed two repeated phrases, fixed the punctuation, and left the joke standing. That is its light-touch gift: intervene, don’t occupy.
Out of curiosity — and because I still had time, which is when all bad decisions begin — I gave ChatGPT the same paragraph. It changed the line to:
“Many parents find it difficult to accept support from their children.”
Accurate. Sensible. Dead.
Another prompt (“preserve the exact joke”) would probably have fixed it — strict constraints are ChatGPT’s home turf. But I didn’t need a negotiation. Claude had left the human being in the paragraph.
This is why I distrust “better for writing,” as if writing were one job. First draft, sensitive edit, subject lines, captions with a character limit — four different jobs wearing one coat. Claude took the first two. ChatGPT took the rest. Although “took” makes it sound tidier than it felt at 9:40 with the clock going.
The subscription I nearly cancelled
I pay for Claude.
Writing is my main work, and Claude saves me the most time where I care most: the first proper draft. I nearly cancelled ChatGPT in February, after two quiet weeks — paying for a second tool just to make lists seemed absurd.
Then a client asked for a month of short-form content: thirty captions, twelve hooks, descriptions in two lengths, all from three articles, all with different calls to action. Claude could have done it — like a novelist asked to fill in parking tickets. ChatGPT produced the volume and mostly followed my banned-word list. It tried to sneak “elevate” past me twice.
So ChatGPT stayed. Claude is the one I trust with a blank page. ChatGPT I trust when the page already has boxes printed on it and somebody needs them filled by Thursday — a different job entirely, and if I’m honest, the one that pays more reliably.
My routine grew out of that division. Angles and outlines start in ChatGPT; the draft itself happens in Claude; then there’s me, deleting anything neither tool had the nerve to question. In practice: three tabs open, one man muttering, “No, the other one,” at his screen.
Where I nearly ruined the piece myself
None of that would have saved the newsletter if I had fed the tools only the service list. My first prompt did: Saturday. Cleaning. Shopping. Companionship. Affordable packages. The results were polite and empty because I had been polite and empty.
The detail that changed it came from the voice note: the milk lasting too long. One real thing. It gave Claude something to write around.
I read the final draft aloud while waiting for the kettle, which takes about as long as a Claude draft and sounds like a small aircraft landing. Anyway. Always do the reading-aloud, even if the neighbours think you host a small radio programme. One sentence — “This service is designed to support families who are navigating the challenges of independent living” — sounded fine on screen and turned to cardboard in my mouth. Cut. The founder’s “another pair of eyes” line became the ending.
I also cut my first paragraph, almost by reflex now — AI drafts warm up in public; humans do too. Paragraph two had the real opening. And I left one section short (price, hours, how to book): some parts do not need to sing. They need to answer the phone.
If you want a test rather than my Tuesday:
“Write 120 words for a beginner who keeps putting off starting a blog. Be encouraging, don’t be cheesy, and give one small action for today.”
Claude will probably give you a person talking; ChatGPT, a framework with better manners. Either way: never publish the raw draft. Raw is where both tools are most average.
Questions people actually ask me
Fine. Which one sounds more human?
Claude. Nine times out of ten, anyway.
Is ChatGPT bad at writing, then?
No — annoyingly, no. Clear, quick, obedient, superb at short-form. Editing too, up to a point: it tidies like a new manager reorganising the whole office when you only asked it to move one chair. One beautiful paragraph, though? I don’t start there. Thirty usable things before a client call? Every time.
Should I pay for both?
The long version, since people keep asking. Only if both earn their keep weekly — mine do, just barely. February nearly killed it: one list in nine days, tea going cold over the cancel button. Then the short-form job landed and it paid for itself in an afternoon. That’s the test — not “is it good?” but “did it do a job this week I’d otherwise have done at midnight?” Two no-weeks running, cancel it. Loyalty is for football teams, not software.
Can readers tell which tool I used?
Mine couldn’t, on Tuesday — they noticed tone, not machinery. The founder rang at 10:31 — seven minutes after it landed, I checked — to say her daughter had read it without rolling her eyes, which in that family is a five-star review. Then: “It doesn’t sound like an advert for a care home.” I’ve had worse feedback dressed as praise. Nobody said the word AI once. Readers spot generic, not brand names.
Last thing
The newsletter went out at 10:24, six minutes early. The client replied: “Sounds like us, finally.”
That is the test I care about now: whether the words sound like the people who have to stand behind them.
Claude usually gets me closer on the first pass. ChatGPT gets the surrounding work done at indecent speed. Then I earn my money on the part neither can do: knowing when “precious moments” has to die.