One Prompt That Makes Any Text Model Write Like a Person

What is "AI flavor"

AI-written text often has a recognizable tone: it loves metaphors, parallel structures and flowery words, it favors "not only... but also..." and "let's go together," and it takes a detour where it should just say the plain thing. It reads like a performance rather than a clear explanation of what happened. This tone is what people call "AI flavor."

This is not a quirk of any one model. Claude, GPT, Gemini, and Chinese models all do it. The good news is that a single prompt can mostly suppress it, and it works with any text model.

One core principle

Anthropic's internal guide for removing AI flavor comes down to one line: if you can say it directly, say it directly. When a plain wording exists, don't reach for metaphors or ornate words, and don't use language that performs the author instead of conveying meaning.

Expanding on that line gives you the prompt below, ready to paste.

A prompt you can use right away

Add this block after your request when you ask a model to write something. Both Chinese and English versions are provided; use whichever you prefer. It works with any text model.

Chinese version:

English version:

Writing rules:
- Say it straight. When a plain statement works, do not reach for metaphor, flourish, or decoration.
- Write to convey meaning, not to perform. Cut anything that shows off the writer.
- Drop filler: no "not just X but Y", no "let's", no "in today's world", no buzzwords.
- One idea per sentence, short sentences first. Use common words.
- No em-dashes for rhythm, no exclamation marks for drama.
- Before finishing, delete any sentence that performs instead of informs.

A comparison

The same idea, before and after adding the prompt:

• With AI flavor: "In this rapidly changing era, AI is not just a tool but a partner fighting alongside you, empowering every person who dares to dream."
• After removing it: "AI is a tool that helps you get work done."

The second sentence loses no information and is more credible. The first looks lively but says nothing more.

How to use it reliably

• Put it last. Place these rules after your specific request so they constrain the final output.
• Pair it with examples. Give the model one or two plain sentences you approve of as a sample; this works better than rules alone.
• Use two steps. Let it write normally first, then send a follow-up like "Rewrite it following the writing rules above and remove all AI flavor." This is often cleaner than getting it right in one pass.
• Don't overcorrect. The rules remove performative tone, not warmth or readability. Necessary transitions, examples and conversational phrasing should stay.
• Do a final check yourself. Models sometimes slip back into the old tone, so read it through and delete any filler that got past the rules.

Further reading

• To make these rules persistent and apply them automatically every time, you can package them as a skill: see Skill Handbook.
• To understand why models write this way, see the LLM Handbook and the Prompt Handbook.