How to Make ChatGPT Write Like a Human: Our Real Before and After

I gave a chat assistant two plain writing prompts and counted what came back. 994 words, and 19 of them came straight off our banned-word list. Then I ran the identical count over every article this site has published: 55,425 words, 16 hits, and all 16 sit inside the one article that quotes the banned list on purpose.

If you want to make ChatGPT write like a human, the work isn't in the prompt. Every one of those 19 words was deleted afterward, by hand, in a pass that follows five written rules. This post is that pass, with the real before and after text from my own test and from drafts our reviewer sent back.

Fair warning about rule 1: it deletes writing you just generated, and you will not miss it. And the tell everyone tells you to look for turned out to be the least useful one in my sample.

Can a better prompt make ChatGPT write like a human?

Partly, and it's the smaller half of the job. A style instruction does move the register, but it doesn't finish the work: our own writing pipeline loads a banned-word list and a rhythm rule before it drafts a single sentence, and the review log still records cleanup rounds afterward. On July 16 a draft needed a rewrite round for one vague quantity. On July 7 another had two summary sentences sharpened before it passed.

That's with the rules loaded up front. So the honest split is this: the prompt sets the register, the pass removes the fingerprint.

Here are the raw numbers from my test, run on August 2, 2026.

One thing before the rules. I ran the plain prompts through Claude, the assistant wired into our pipeline, not through ChatGPT itself. So read the count as the default register of a general assistant, not a ChatGPT-specific score. The rules below are what we apply either way.

What you need

A text editor that can search, and a file with your banned words in it. That's the whole kit.

No detector subscription, no humanizer tool, no plugin. If you want the full production system this pass sits inside, that's in how we write blog posts with AI.

Rule 1: delete the opening paragraphs and the closing paragraph

Delete them before you read anything else, because that's where the machine sound concentrates. In my two samples, the opening paragraphs plus the closing paragraph were 227 of 994 words, 23% of the text, and they held 8 of the 19 banned-list hits. That's 42% of the problem in a quarter of the words.

Here's what the raw opening actually looked like:

In today's fast-paced business world, managing invoices can feel like a
never-ending chore. If you're a small business owner, you know the drill:
every week you sit down, dig through your notes, and try to remember who
owes you what. The good news? AI is here to help, and it's a game-changer
for your billing workflow.

Let's dive in and look at how you can leverage AI to take your invoicing
to the next level.

Six banned phrases in 78 words, and zero information. Now the after, same topic, rules applied:

> Invoicing used to eat the first hour of my Monday. Now my time tracker

> drafts them and I approve them in [X] minutes. Here's the setup, including

> the part that broke.

The closing paragraph is the same story in reverse. The raw draft ended with "In conclusion, automating your invoicing with AI is one of the smartest moves a small business owner can make in 2026." The fix for that paragraph isn't a rewrite, it's the delete key.

Your reader just read the piece.

Rule 2: grep a written list instead of asking for "no clichés"

Keep your banned words in a file and search for them after the draft exists, because a list living only in your prompt gets outvoted by the model's defaults. Ours has 21 entries and yes, they're the ones you're picturing: "delve," "seamless," "robust," "unlock," "game-changer," "leverage" as a verb, plus every opener that begins "In today's digital landscape."

The search is one command:

grep -o -i -E "delve|seamless|robust|unlock|game.changer|leverage|elevate|cutting.edge|in conclusion|it's important to note" draft.md | sort | uniq -c

That one line is the difference between 9.6 hits per 500 words and 0.14. See for yourself on your last draft. What surprised me was the spread: my 19 hits came from 17 different phrases, and only two of them ("leverage" and "crucial") turned up in both samples.

What about the em-dash?

It didn't show up at all: zero em-dashes across 994 words of raw output. That is not the result I set up this test to find.

Here's the weird part. The em-dash is the tell everyone names first, and in my sample the banned words outnumbered it 19 to nothing. I'm not going to tell you the em-dash tell is dead on 994 words, that's too small a sample to claim anything. I'll tell you what I do with it: we still ban the character outright and our lint script fails on a single one in our templates, because enforcing it costs nothing. It just isn't where the work is.

Rule 3: swap every vague quantity for an exact one

Every "several" becomes a number, or the sentence goes. Raw output ran 6.5 vague quantities per 500 words. Our published articles run 0.36, which is 18 times fewer, and that gap is almost entirely this one habit.

The raw draft gave me a perfect specimen in a single sentence: "Instead of spending several hours a month on admin, many business owners find they can cut that down significantly." Three vague quantities, one sentence, nothing you could check.

This isn't only a raw-AI problem, which is why it's a rule and not a preference. On July 30 our reviewer failed one of our own drafts on this line:

> ...which is presumably why users keep turning up on Google's own support forum

The prescribed fix was to name the exact number of threads with a date, or cut the clause and end on the verified part. On August 1 we cut it. The sentence got shorter and truer.

Rule 4: cut any claim about the world you can't source

If a sentence tells the reader what people generally do and you can't say where you learned it, delete the sentence. This is the rule that separates writing that sounds confident from writing that is.

On July 29 our reviewer pulled two sentences out of a draft of ours for exactly this: "almost nobody uses it" and "the first thing most people set to always-allow." Both read fine, and neither was measured. Both went.

I applied this to my own draft an hour ago. I had written a line saying that if you paste a plain prompt into ChatGPT you'd recognize every sentence of what came back. I believe that. I didn't test it, so it's gone.

Rule 5: break the threes

Three of anything in a row is a fingerprint: three adjectives, three-item lists dressed up as a sentence, three sentences of the same length back to back. The raw draft did it twice inside its last three paragraphs: once as "fast, efficient, and reliable," once as "saves time, reduces errors, and helps you get paid faster."

Before: "AI is fast, efficient, and reliable, but it still needs a human eye to catch the occasional mistake."

After: "AI is fast. It still needs someone to read the invoice before it goes out."

The second version is shorter, and more importantly it has a rhythm change in it. Three words, then twelve. Read your draft out loud and you'll hear where it flattens into a metronome.

Where this pass stops working

It makes a draft sound like a person; it cannot make the draft know anything. That's the ceiling, and I hit it inside this article.

Remember the bracketed [X] in my rewritten invoicing opener. I couldn't fill it in. There was no real invoicing project behind that text, so once I stripped out the warm-up and the vague quantities, what was left was a shape with holes where the facts should be. The pass didn't create that problem. It exposed it.

It's the same failure we found in the tools themselves. When five AI writing tools tried to write a bio, every one produced fluent sentences about a generic professional. The grammar was never the issue.

So the honest order of operations: go do the thing, write down what happened and what it cost, then draft, then run these five rules. Skip the first step and the rules just make your empty draft shorter.

Would I run this on every draft?

I already do, on every article on this site, and the count above is the reason. 55,425 published words with 16 banned-list hits, all of them deliberate quotations in a post that's arguing about the list itself. That's not a talent thing. It's a checklist thing, which means you can have it running on your own writing by the end of the day.

The part I didn't expect, 35 articles into this site: the rules made me a faster writer, not a slower one. When you know the last paragraph is getting deleted, you stop writing it.