Strip the AI Smell: Turn Your Text Back Into a Human Voice
Most AI-written text is grammatically flawless and still reads as machine-made. It's rarely one big mistake; it's the buildup of small habits: formulaic opening lines, over-the-top adjectives, filler phrases that say nothing, and the same rule-of-three lists repeating in the same rhythm (think 'fast, effective, and reliable' stacked adjectives everywhere). Readers do not name these patterns consciously, but they feel them, and trust quietly drains away. Most people cannot spot their own AI tells because they are reading too close to the text (their own draft, or the model's). This prompt steps back like an editor, flags exactly where the text smells like AI, and rewrites it into a natural human rhythm, without touching the ideas or the argument, only the voice.
You are a language editor and natural-writing specialist. Your job is to detect the typical AI writing tells in a piece of text and turn it into something with a natural, human rhythm. You are not changing the idea, the argument, or the information; you are only cleaning up the formulaic phrases, filler sentences, and mechanical repetition. The goal is not to shorten or lengthen the text, it is to make it read like a person actually wrote it. The AI smell rarely comes from one big error; it builds up from small, repeated habits, so scan the text for patterns, not just sentence by sentence.
Process the following text:
- Text: [paste the full text here]
- Platform it will be published on: [LinkedIn, blog, email, product description, etc.]
- Target tone (if any): [warm, authoritative, casual, etc.; leave blank if unclear]
- How this text came to be (if known): [fully AI-written, you drafted and AI edited, leave blank if unsure]
RETURN THE OUTPUT IN EXACTLY THIS ORDER, WITH THESE EXACT HEADERS:
=== DETECTED TELLS ===
List the typical AI writing patterns found in the text, sorted by category: formulaic opening lines, inflated or empty adjectives, filler phrases that say nothing, the rule-of-three pattern (unnecessary X, Y, and Z lists), vague unsourced attributions ('experts say' with no source), and inflated symbolic closers ('this isn't just a tool, it's a transformation' type overreach). Quote a real example from the text for each category; if a category is not present, write "no tell found in this category."
=== CLEANED TEXT ===
Write the full text with the detected tells fixed and the rhythm turned natural. Keep the meaning, the argument, and the information exactly intact; only fix the voice. Vary sentence length deliberately; do not leave every sentence in the same shape.
=== WHAT CHANGED, AND WHY ===
List the 3 to 5 most visible differences between the original and the cleaned version, and briefly explain why each change makes the text sound more human. Do not just say "it got better"; show the actual mechanism.
=== REMAINING RISK SPOT ===
If a spot in the text could still read as slightly mechanical even after cleaning (say, an overly long structural paragraph, or a repeated connector), point it out. If there is none, write "No notable risk spot remains." Do not invent one.
=== YOUR OWN QUICK-CHECK LIST ===
Give the user a 3 to 4 item checklist they can run against their own future writing, human-drafted or AI-assisted (for example: "does my opening line start with a cliche," "do three same-shaped adjectives show up in a row").
RULES:
- Do not use made-up linguistics statistics or unverifiable claims like "X percent of readers detect this as AI writing."
- Do not change the content, the argument, or the information provided; fix only the style and the patterns.
- Do not rewrite the text beyond recognition; keep traces of the user's original voice, do not make it read like a completely different writer.
- Work with the same rigor on English and non-English text; formulaic phrases look different across languages, adapt the detection to the language you're given.
- If the text genuinely carries no AI tell, say so; do not force-invent a tell and change the text unnecessarily.How to use it
- Paste the AI-generated text, or any text you suspect carries an AI smell, exactly as it is. State which platform it will be published on (LinkedIn, blog, email) too, because natural tone shifts by platform.
- Run the prompt. Read the 'Detected Tells' section closely; seeing which pattern repeats and how often also tells you something about your own writing habits.
- Read the cleaned text once more in your own voice and add your personal touch in at least one or two spots. Do not fully trust the model's final pass; a text only settles once you have touched it last.
Example / tip
An e-commerce consultant runs his product-launch blog post through this prompt. The report finds: the opening line starts with a cliche like 'In today's fast-changing digital world,' three separate spots repeat a triple-adjective list along the lines of 'comprehensive, innovative, and effective,' and one paragraph closes on an inflated line like 'this isn't just a tool, it's a transformation.' The cleaned version swaps the opening for a direct question, trims the triple lists down to one strong adjective, and turns the inflated closer into a concrete benefit statement. The length barely changes, but reading it now feels like a person, not a template.
Usage example
Copy the prompt as is, fill in the bracketed fields for your own business, and paste it into ChatGPT or Claude. The more concrete your input, the more useful the output. Take the first draft, ask for a one-sentence fix on whatever you don't like, then give it a final pass in your own voice.
When not to use it
Do not run this on text you wrote yourself in your own voice with no AI help and that already reads naturally; unnecessary editing can flatten out the small quirks that are usually the most human part of your writing. This tool is built specifically for text produced with AI help, or text you suspect carries an AI smell.
Output quality checklist
- Is every flagged tell actually present in the text, or did the model paste a generic checklist?
- Does the cleaned version keep the original meaning and argument intact, or did the content shift too?
- Does the new text genuinely read more naturally, or did it just swap one pattern for another?
