speak-human-tw（說人話）是个给繁体中文去 AI 味的改写 skill，能抓出 35 种以上的 AI 写作套路，顺手把中国用语和半形标点改成台湾写法…

On July 12, 2026, developer Geek released a rewriting skill called ‘speak-human-tw (說人話)’ on platform X. The tool targets traditional Chinese (Taiwan) users, aiming to address common ‘AI flavor’ issues in AI-generated text, including mechanical sentence structures, repeated specific vocabulary, and overuse of connecting words.

According to the developer’s description, the tool can identify over 35 AI writing patterns and automatically perform two types of conversions: one is to replace common mainland Chinese terms with Taiwanese equivalents (e.g., ‘通過’ → ‘透過’, ‘信息’ → ‘資訊’); the other is to correct half-width punctuation to full-width (e.g., English commas to Chinese commas). These adjustments are believed to make the text read more naturally and in line with local Taiwanese writing habits.

Although the developer has not released complete technical documentation, based on functional descriptions, the tool likely uses pattern matching or a lightweight rules engine to flag common AI features. For example, it might detect high-frequency routine phrases like ‘值得注意的是’ (It is worth noting) or ‘總而言之’ (In conclusion), or analyze sentence length distribution anomalies.

The conversion part relies on predefined dictionary mapping tables and punctuation rules. This approach has the advantage of being lightweight and runnable offline, but the downside is difficulty handling context-dependent vocabulary choices (e.g., ‘質量’ in Taiwanese can refer to product quality or physical mass, requiring disambiguation).

speak-human-tw is designed to integrate into development environments like Claude Code, Codex, and Cursor, suggesting its primary target users are programmers or technical writers who frequently use AI to generate comments, documentation, or blog content but want the output to adhere to Taiwanese usage conventions.

However, the tool is currently limited to specific editor ecosystems; non-developers need to manually convert texts. Additionally, whether the coverage of 35 patterns is sufficient to handle output styles from different AI models (e.g., GPT-4, Claude 3.5, Gemini) remains to be verified.

The post quickly garnered about 3,900 views and sparked 9 comments, indicating a demand for such tools in the traditional Chinese community. Some users expressed delight at having a Taiwan-specific AI post-processing tool, but others questioned its actual effectiveness: ‘Do the 35 patterns cover all common modes?’ ‘Will it over-correct, e.g., converting intentional mainland Chinese terms as well?’

Due to a lack of independent evaluations, the tool’s recall and precision are currently unverifiable. The developer only posted on a community platform without providing detailed test data or samples, limiting confidence.

The emergence of speak-human-tw highlights the importance of localizing AI text. Global AI models are mostly trained on English or simplified Chinese data, and when output is used in traditional Chinese environments, they often carry accents or terminology biases. Taiwanese users either accept these biases or manually correct them—the latter being time-consuming and inconsistent.

This tool represents a decentralized solution: the community develops lightweight rules to correct model output. This may drive more similar projects (e.g., for Hong Kong Cantonese, Malaysian Chinese), but also reminds model developers to pay more attention to balanced training data for language varieties.