Write with AI without sounding like AI

The problem is not that AI writes badly

The problem is that AI writes the same. Every assistant, left to its defaults, produces prose with the same rhythm, the same hedges, the same tidy three-item lists and the same closing line that sums up what you just read. Readers have learned the pattern in two years, and they skim past it the way they skim past stock photos. Being skimmed is the same as not being read.

The fix is not to stop using AI. It is to stop letting it choose. A model is a strong drafting engine and a weak editor of its own work. So you split the job into stages and keep the decisions in the stages that need a human.

Stage one: the brief

A bad brief is "write an article about AI note-taking apps". A good brief is a page long and takes twenty minutes, which is why most people skip it and get the article they deserve.

Your brief contains:

• The reader, in one sentence, with the situation they are in. "A freelance designer who has tried three note apps and abandoned all of them."
• The one thing they should be able to do after reading. Not "understand"; do.
• Your position. Every good piece has one. "Most note apps fail because they optimise for capture, not retrieval."
• Three to five facts you have already verified, with sources, and a rule that nothing else may be stated as fact.
• What you will not cover, so the draft does not sprawl.
• The quotes you collected in step one, in the reader's own words.
• Length, structure, and two examples of writing you admire, pasted in.

Paste that into Claude or ChatGPT. The difference between a one-line prompt and a real brief is bigger than the difference between any two models.

Stage two: the draft

Ask for a draft, then ask for it again, differently. Two useful moves:

• Request two drafts with different structures: one that opens with the reader's problem, one that opens with your position. Pick the stronger opening and keep going from there.
• Ask the model to mark every sentence it is not confident is true with a bracketed note. It will comply, and you will be surprised how many brackets appear.

Do not accept the draft's headings. Models default to the same five section names. Write your own, in the words the reader would use to search.

Stage three: your voice

This is the stage that separates a site people return to from one they do not. It cannot be delegated, but it can be made fast.

Read the draft aloud. Anywhere you would not say it, change it. Add the sentence only you could write: the thing you noticed, the mistake you made, the number from your own account. One specific line per section that could only have come from you is enough to make the whole piece read as human, because it is.

Then do the deletions. These are the slop patterns worth hunting, and the reasons they hurt:

• Announcements of what is coming: "In this article, we will explore." The reader knows. Cut the paragraph.
• Summaries of what just happened: "In conclusion, as we have seen." Cut.
• Empty intensifiers and throat-clearing: "truly", "incredibly", "it should be noted". Delete the words; the sentence survives.
• The tidy triad: three adjectives where one would do. "Fast, simple and powerful." Pick one.
• Hedge stacks: "may potentially help to". Decide: it does or it does not.
• Rhetorical questions that set up an obvious answer. State the answer.
• The reassurance close: "With the right tools, anyone can." Nobody has ever been reassured by that sentence.
• Every metaphor about journeys, unlocking, or navigating. The reader has met them a thousand times this month.
• The vocabulary that now marks a text as machine-written and triggers an instant skim. We keep a banned list and a search-and-delete habit. Yours will differ. Build it from the words you keep seeing.
• Paragraphs of identical length. Vary them. A one-line paragraph is a tool.

You can ask the model to do the first pass of these deletions. Tell it: "Remove every sentence that announces, summarises or reassures. Remove filler intensifiers. Replace each abstract noun with a concrete one where possible." Then read it again yourself, because it will have missed some and introduced one or two new ones.

Stage four: the fact-check

An AI draft states things with the same confidence whether they are true or invented. So the check is mechanical, not a feeling:

• Every number, date, price, name and quote gets a source, or gets cut. No exceptions for things that sound right.
• Prices and thresholds get checked against the vendor's own page on the day you publish, and dated in the text where they matter ("as of September 2026").
• Perplexity is useful for a fast pass because it links sources, but the link is the check, not the summary. Open it.
• Quotes are only quotes if you can point to where they were said. Never let a model produce one. If a draft contains a quotation you did not paste in, it is invented.
• Anything legal, medical or financial gets the same treatment plus a line that says it is not advice.

Keep a short sources list at the bottom of your working document even if you do not publish it. When a reader writes in to say a number is wrong, and they will, you can answer in a minute.

Labelling

If a page is substantially AI-generated and published without real editorial oversight, the EU's transparency rules now expect it to be labelled; step five covers this in more detail. A short, honest line about how you work, on your about page and in your footer, costs nothing and reads as confidence. Ours says that we draft with AI, edit by hand and check every number. Then we have to actually do it.

The workflow in one line

Brief for twenty minutes, draft in five, edit for an hour, check for thirty minutes. Two hours for a piece that people will finish. That is faster than writing alone and better than publishing raw.

Checklist

• Brief written with reader, action, position and verified facts before any drafting
• Two drafts generated with different openings and the stronger one chosen
• Draft read aloud, one specific human line added per section, slop patterns deleted
• Every number, name and quote traced to a source or removed
• A plain line about how the piece was made added where readers can see it