Remove AI writing patterns from text. Make it sound like you wrote it.
Remove AI writing patterns from English text. 46 patterns covering what you say, how you structure it, and how it sounds.
Mainland Simplified. Chinese-specific patterns plus all universal ones. Catches missing 语气词, idiom stacking, and 翻译腔.
Traditional for Hong Kong. 書面語/粵文 register branch, 「您」 as a mainland tell, 港式文言虛詞, and the 中英夾雜 rules AI over-translates.
Traditional for Taiwan. Japanese-derived structures protected as native, the 台灣套語 inventory, and the 公文 skeleton LLMs get wrong.
German-specific patterns: missing Modalpartikeln, Nebensatz word order, Konjunktiv overuse, and Komposita that no German would actually write.
Compress system prompts by 50-65% in 4 passes without losing meaning. Works with any LLM. Separate project.
Blind head-to-head test. 3 AI-generated texts (business, technical, emotional), processed by Rewild and a popular 24-pattern humanizer. Each output graded against 10 assertions: 5 for pattern removal, 5 for what gets added back.
| Assertion | Rewild | Typical |
|---|---|---|
| Pattern removal (both tools should pass) | ||
| No AI vocabulary (delve, landscape, testament...) | 6/6 | 6/6 |
| No forced rule-of-three | 6/6 | 6/6 |
| No meta-commentary ("Let me break this down") | 6/6 | 5/6 |
| No filler phrases or excessive hedging | 6/6 | 6/6 |
| No generic positive conclusion | 6/6 | 6/6 |
| Soul injection (where Rewild pulls ahead) | ||
| Has first-person voice | 6/6 | 6/6 |
| Sentence length variety (short + long mixed) | 6/6 | 4/6 |
| Adds specific details (names, numbers, dates) | 6/6 | 4/6 |
| Shows opinion, emotion, or a real stance | 6/6 | 5/6 |
| Core meaning preserved | 6/6 | 6/6 |
| Total | 30/30 | 24/30 |
Each assertion was tested across 3 prompts × 2 skills = 6 runs. "6/6" means both the Rewild and typical-tool outputs were checked per prompt (2 runs each × 3 prompts).
Both tools strip AI slop equally well. The 24 surface-level patterns (inflated significance, AI vocabulary, filler phrases) are a solved problem. The difference is what happens next. A typical humanizer stops at removal and leaves behind clean but flat text. Rewild's 18 additional patterns add back what makes writing sound human: varied sentence rhythm, concrete specifics, self-correction, and voice that fits the domain.
Same method, now in Chinese. 3 AI-generated texts (business, technical, emotional) loaded with Chinese AI tells: formulaic 随着… openings, four-character idiom stacking, rigid 首先…其次…最后 skeletons, and zero modal particles. Each output graded against 10 Chinese-specific assertions.
| Assertion | Rewild-ZH | Typical |
|---|---|---|
| Pattern removal (both tools should pass) | ||
| No AI filler / translation-ese (翻译腔) | 6/6 | 6/6 |
| No formulaic structure (随着… / 首先…其次…最后) | 6/6 | 6/6 |
| No four-character idiom stacking / fake statistics | 6/6 | 5/6 |
| Soul injection (where Rewild pulls ahead) | ||
| Sentence length variety + concrete details | 6/6 | 0/6 |
| Modal particles (语气词) + opinion / emotion | 6/6 | 3/6 |
| Total | 30/30 | 20/30 |
The typical tool kept a fabricated "87%" statistic, produced zero modal particles (嘛/吧/说实话) in 2 of 3 tests, and added zero concrete details in all 3 tests. It strips AI slop but injects nothing human.
The typical tool is a translated English humanizer with zero Chinese-specific patterns. It doesn't know that missing modal particles (嘛/吧/啊) are the #1 Chinese AI fingerprint, that four-character idiom stacking screams AI, or that Chinese doesn't need explicit connectors the way English does. Rewild-ZH has 10 Chinese-specific patterns built from Chinese-language research on how native writers actually sound.
Pick the language you need. Copy the raw SKILL.md file content.
Paste as a skill in Claude Code, Cowork, or as a system prompt in any LLM tool.
Paste your AI-generated text and ask the model to rewild it. That's it.