How to clean Claude or GPT output before publishing

A concrete workflow for taking model output to publishable text without losing what it actually says.

Guides · 2 min read

If you use a model as a drafting tool and publish the result under your own name, there is a short list of things worth doing every time. None of it is about evading anything — it is about the output not looking like output.

1. Paste, do not retype

Everything in step 2 disappears if you retype. Paste the raw output.

2. Strip the characters first

Model output reliably contains:

• Em dashes, used heavily
• Curly quotes and apostrophes rather than straight ASCII
• The single-codepoint ellipsis U+2026
• Occasional non-breaking spaces and other exotic whitespace
• Sometimes zero-width characters, usually inherited from formatting in context

This layer is mechanical. Removing it is lossless, needs no model, and is verifiable by re-scanning: the count should hit zero.

Do it first, before any rewriting, so nothing has a chance to reintroduce it.

3. Cut the scaffolding

Delete rather than rewrite:

• Sentence-opening Moreover, Furthermore, Additionally, In conclusion
• It is important to note that, it is worth remembering
• In today's fast-paced world and its relatives

The claim underneath survives. These phrases carry no information — that is why they are a tell.

4. Replace the inflated vocabulary

delve to look at. leverage to use. robust to strong or reliable, unless it is the statistical term. tapestry, landscape, realm to the actual subject.

One-for-one substitution, keeping the sentence otherwise intact.

5. Break the rhythm

This is the step that matters most and gets skipped most. Model prose runs sentence after sentence of near-identical length, and that evenness is more recognisable than any individual word.

Split one long sentence. Join two short ones. Let one run three words. You do not need to do this everywhere — one intervention per paragraph changes how the whole thing reads.

6. Check the facts

Non-negotiable, and nothing about cleanup touches it. Verify every figure, name, date, quotation and citation independently. Cleanup makes text read as yours; it does not make it true.

7. Read it aloud

The fastest test available. Machine-written prose reads as relentlessly even. If it sounds like a brochure, the rhythm work is not done.

What to keep

Do not sand off everything. If the draft says something clearly, leave it clear. The goal is text that reads as written by a person, not text that reads as deliberately roughened. Over-editing is its own tell.

Run a free scan on your draft and you will see exactly which characters and phrases are present before you start.

Related: why AI text is full of em dashes · the words that give away AI writing