deai（中文写作降AI率）

Chinese LLM outputs often suffer from "AI flavor"—predictable structures, corporate jargon, and excessive transition phrases. This skill implements a multi-layered pipeline to strip these artifacts from web novels, social media posts, and professional articles. It integrates five distinct de-AI frameworks to address issues ranging from high-level narrative architecture to character-level word bans.

What it does

• Multi-level diagnostic identifies if text is fiction, non-fiction, or mixed to apply relevant stylistic rules.
• Structural de-sloping flags narrative issues like "fake profoundness" or repetitive paragraph openers before editing.
• Phrase-level filtering targets specific "AI-isms" including 17 variants of straw-man arguments and empty transition phrases.
• Hard-word banning utilizes a script-based approach to clear blacklisted jargon and excessive punctuation patterns.
• Information preservation ensures no new facts are invented and core plot points or data remain intact during cleaning.

How it works

1. Pre-check: The agent analyzes the input text, identifies the genre, and selects a processing mode: Clear, Reconstruct, Detect, or Write.
2. Structural Analysis: For fiction or reconstruction tasks, it diagnoses narrative skeleton issues and proposes 2-5 specific actions.
3. Pass-based Editing: It executes a series of passes—first cutting filler, then applying sentence-level gate rules, and finally clearing banned words.
4. Cold Reading: The agent performs a final "human-like" review to ensure the text doesn't become too thin or telegraphic.

Frameworks & tools

The skill synthesizes rules from human-writing, lieflat, sepia, oh-story, and qifeng. It utilizes a Python-based check_deai.py script for quantitative validation of banned terms and stylistic markers.

Why this beats prompting it yourself

Generic prompts often result in "over-cleaning," where the AI removes so much detail that the text develops a new, equally recognizable "over-edited" fingerprint. This skill uses a graded arbitration system that prioritizes author style over general rules, ensuring the output remains natural rather than just robotic in a different way.

Use cases

• Stripping "model-speak" from Zhihu or WeChat official account drafts.
• Cleaning web novel chapters to remove repetitive AI-generated narrative beats.
• Converting corporate-style AI summaries into natural, spoken-feeling Chinese.
• Auditing existing text to flag specific AI-triggered markers without modifying the content.

Known limitations

This skill focuses on removing AI-style artifacts and is not a general grammar checker or proofreader. It requires manual confirmation before making significant structural changes to fiction.