Raymondhou0917/speak-human-tw

What it solves

This project addresses the "AI flavor" (AI-generated feel) in Traditional Chinese text, where content is grammatically correct but feels unnatural or robotic. It specifically targets the removal of common AI writing patterns, the correction of Mainland Chinese terminology into Taiwan-localized Traditional Chinese, and the fixing of punctuation errors common in AI outputs.

How it works

It functions as a "skill" (a set of instructions and rules) for AI agents like Claude Code, Cursor, and Codex. Instead of generating content from scratch, it acts as a a proofreading tool that follows a six-step process:

1. Contextual Analysis: Determines the tone based on the scenario (e.g., social media, newsletters, sales pages).
2. Protection: Locks essential facts (prices, links, names) to prevent accidental changes.
3. Scope Definition: Ensures long-form content is not arbitrarily shortened.
4. Categorized Rewriting: Applies rules to remove 35+ types of AI traces (e.g., exaggerated meanings, formulaic openings, emoji bombing).
5. Verification: Cross-checks against the protection list to ensure factual integrity.
6. Self-Evaluation: Scores long-form text on a 50-point scale to ensure quality.

By default, it operates in a two-round interactive mode: first providing a numbered list of suggested changes and their reasons, then applying the changes only after user confirmation.

Who it’s for

Content creators, marketers, and office workers who use AI to draft content in Traditional Chinese for newsletters, social media posts, sales pages, and professional emails.

Highlights

• 35+ AI Trace Detection: Identifies specific robotic patterns like "not A but B" stacks and formulaic conclusions.
• Taiwan Localization: Includes a comprehensive mapping of Mainland Chinese terms to Taiwan-specific usage and punctuation rules.
• **Interactive Proofreading: Prevents accidental overwriting of original drafts by requiring user approval for specific edits.
• Fact-Preservation: Uses a protection list to ensure critical data and links remain untouched during the rewriting process.
• Benchmark Testing: Includes a 42-case benchmark to ensure necessary changes are made without "over-killing" (mis-editing) factual or legal text.