ChatGPT vs Claude for long-form writing: one brief, two drafts, an honest edit

Short answer: for a single long piece with a distinct voice, Claude needed less editing. For a piece that needed research done first and a structure found from scratch, ChatGPT got further faster. Both produced drafts we would not publish without a second pass — which is rather the point of this site.

The rest of this article is how we got there, so you can decide whether our brief looks like your work.

The brief

A 2,000-word reported feature on why small manufacturers are bringing production back onshore. We supplied:

Six pages of interview notes from three named sources (all fictional, all internally consistent)
Two paragraphs of a previous article by the same writer, as a voice sample
A hard rule: no statistics that are not in the notes, and no invented quotes
A target reader: a business owner, not an economist

Both tools got the identical prompt, the same notes, and the same follow-up instructions. We ran each three times and edited the best draft.

First draft: structure

ChatGPT’s draft had the better shape. It opened on a specific source, moved to the wider trend, gave the counter-argument its own section, and ended with a practical takeaway for the reader. It read like a magazine feature.

Claude’s draft was structurally more conservative — trend first, then the sources as evidence — but it was the only one of the two that used the interview notes in the order that made the argument strongest rather than the order they appeared in the file.

Call it a narrow win for ChatGPT on structure.

First draft: voice

This is where the gap opened. The voice sample was dry, slightly wry, short sentences. ChatGPT picked up the short sentences and dropped the wryness after about 400 words, settling into its house style: balanced sentences, a summary line at the end of every paragraph, “it’s worth noting” twice.

Claude kept closer to the sample through the full piece. Not perfectly — it over-indexed on the wryness in two places and we cut a joke — but an editor who knew the writer would have believed the draft was theirs after a light pass. That was not true of the ChatGPT draft.

First draft: factual discipline

Both obeyed the no-invented-statistics rule. ChatGPT added one hedge that implied a statistic (“a growing share of manufacturers”) that was not supported by the notes; it was not a fabrication, but it was the kind of line that gets a “citation?” from a good editor. Claude stayed inside the notes.

Neither invented a quote. That would not have been true two years ago.

The edit

We timed the edit to a publishable state, by the same editor, on separate days, blind to which was which.

ChatGPT draft: 54 minutes. Most of the time went on voice — cutting summary sentences, breaking up balanced constructions, removing hedges. Structure was largely untouched.
Claude draft: 31 minutes. Most of the time went on moving one section earlier and cutting two lines that were trying too hard.

Round two: asking for changes

The follow-up test is the one that matters for real work. We asked each tool to tighten the third section by about 150 words without touching anything else.

Claude returned the third section, shorter, and left the rest alone. ChatGPT returned a full new draft; the third section was shorter, but sentences elsewhere had also changed, including one we had already fixed. This behaviour is consistent across the tests we ran for this site, and it is the single biggest reason long-form writers in our group have drifted to Claude: it edits the way an editor does.

Where ChatGPT wins outright

Two scenarios where we would pick ChatGPT for long-form:

You have no notes yet. ChatGPT’s research and outlining loop is faster. Ask for a research plan, follow it, feed the results back in — the tooling around that loop is more mature.
You need three different angles before you commit. Ask both for alternative structures and ChatGPT’s are more varied. Claude’s alternatives tend to be variations on one shape.

What this means for you

If your writing has a voice you care about and you edit your own work, start with Claude and use ChatGPT for the research phase. If you write a lot of different things for different audiences and care more about speed than signature, ChatGPT as a default is defensible.

Either way, budget for the second draft. That is not a criticism of the tools. It is what they are for.