You can spot AI copy from across the room: the throat-clearing intro, the three-item list that didn't need to be a list, the word "delve," the fake intimacy, the em dash addiction, the closer that says nothing.
The fix is not "write better prompts." The fix is constraints plus samples.
The 60-second install
Paste this into custom instructions (Claude, ChatGPT, Gemini — all of them):
Write like a specific person, not a brand. - Short sentences. One idea each. - No: "I'd be happy to", "In today's fast-paced", "It's important to note", "delve", "landscape", "unlock", "game-changer". - No emojis unless I ask. No hashtags. - Don't open with a question. Don't close with "Let me know if you need anything else." - Prefer verbs. Cut adjectives that don't change the meaning. - If a sentence could appear on a motivational poster, delete it.
Then give it three humans
Models imitate. Give them something worth imitating. Three emails or posts you actually published. Label them "match this." One piece you hated. Label it "never this."
When a draft still sounds like a model, don't rewrite the whole thing. Highlight the worst paragraph and say: "This paragraph is performing. Make it sound like sample 2. Keep the facts."
The truth pass
Before anything goes public, I run one extra turn:
Mark every claim as KNOW / INFER / GUESS. Delete GUESS. Soften INFER. Keep KNOW. Score your confidence 0–100. If under 80, tell me what would raise it.
That pass is how you stop publishing confident nonsense. It also makes the remaining sentences shorter, which is most of what "human" actually means.
If you want the longer version of this as a workflow — critique, then rewrite, then changelog — steal the "Critique Before You Read It" prompt on the library page.