A prompt that teaches Gemini to write less like Gemini

What it does

This repository contains one elaborate system prompt that instructs Gemini 2.5 Pro to act as a humanizing editor for AI-generated Chinese text. The prompt directs the model to diagnose mechanical tics—stiff transitions, deadened emotion, and robotic stock phrases like 值得注意的是—then rewrite the piece with varied rhythm, sensory detail, and implicit connective tissue. It outputs a structured analysis followed by the full polished article.

The interesting bit

The author treats “AI flavor” as a measurable contaminant rather than a vague annoyance, citing specific reductions on the Zhuque detection model: a 5000-word rewrite reportedly drops from roughly 70% AI-likeness to about 17%. The entire product is the prompt itself—no code, no wrapper, no API—just a carefully taxonomized set of constraints that forbid numbered transitions and demand show-don’t-tell.

Key highlights

• Explicitly targets reasoning models; only tested on Gemini 2.5 Pro
• Claims measurable results on the Zhuque AI detector: ~70% AI flavor down to ~17% on 5000-word rewrites
• Expands text (1000 to 2000 words) while reportedly keeping AI-flavor increase below 22%
• Structured output requires the model to first diagnose flaws, list strategies, and then deliver the rewrite
• Includes granular style rules: avoid passive voice, kill listicle transitions, balance colloquialism with register

Caveats

• Only validated on Gemini 2.5 Pro; transferability to other models is untested
• Effectiveness is benchmarked against the Zhuque detector, which may not correlate with actual human perception
• The repository is essentially a prompt template, not a standalone tool or library

Verdict

Grab it if you need to clean up long-form Chinese content for AI-detection scanners or simply want less robotic copy. Skip it if you were looking for a model-agnostic library or automated pipeline.