If you draft with ChatGPT, Claude or Gemini, the draft comes back with the same habits every time. Readers have learned to spot them. This checklist is the order we'd edit in, from the fixes that matter most to the cosmetic ones.
The goal is writing people actually want to read: specific, direct, and in your own voice. Fooling an AI detector is beside the point.
Delete the first sentence if it restates the question, praises it, or announces what's coming ("Great question!", "Let's dive in", "In this article, we'll explore"). Then look at the last paragraph. If it starts with "In conclusion", "Ultimately" or "At the end of the day", or ends on a short line meant to sound deep, cut it. Most drafts get better from these two cuts alone.
Guides: chatbot leftovers, staged run-ups, fake-profound endings.
Search the draft for words that signal importance without proving it: "pivotal", "crucial", "transformative", "game-changer", "revolutionize". For each one, ask "how much?" or "compared to what?" and write that instead. If you don't know, the claim probably shouldn't be there.
Guides: inflated significance, overused AI words, full word list.
"It's not about X, it's about Y." "This isn't just a tool, it's a movement." Models love this shape because it sounds insightful. Usually nobody argued X in the first place. State Y and give the reason.
Guides: not X, but Y, strawman openers.
AI drafts tend to come in threes (fast, simple, and powerful), stack short fragments for drama (No fluff. No jargon. Just results.), and lean on em dashes. Merge the fragments into one sentence with a real claim. Keep a list of three only if there really are three things. Swap most dashes for periods or commas.
Guides: rule of three, rows of fragments, em-dash overuse.
Chat tools format everything as bullets with bold labels and emoji. That's fine for a reference doc. In an email, post or essay, turn the bullets back into sentences. If the bold label just repeats what the bullet says, delete it.
Guides: bold-label lists, emoji decoration.
A number from your own work, a customer's exact words, a mistake you made, the tool you actually use. Models can't invent these without lying, so they're the fastest way to make a piece sound like a person wrote it. One concrete detail per section is a good target.
Anything you'd never say to a colleague, rewrite until you would. "Moreover" and "furthermore" usually go first.
Paste the result into the SlopScore checker. It highlights each remaining pattern and says why. Under 20 usually reads as human. Don't chase zero: a few dashes and lists are normal, and SlopScore ignores weak signals on their own for that reason.