A style guide disguised as a Claude skill

What it does

Humanizer-zh is a Claude Code skill that rewrites AI-generated Chinese text to sound human. It ships as a translated and adapted version of blader/humanizer, with extra tooling borrowed from stop-slop. You invoke it with /humanizer-zh and paste your slop; it returns prose with the AI fingerprints removed.

The interesting bit

The project treats AI writing as a pattern language rather than a vibe. It documents 24 specific tics—“negation parallelism,” “fake scope,” “three-part rule overuse,” “obsequious tone”—and trains Claude to spot and excise them. The Chinese adaptation is nontrivial: some English patterns (title case, em-dash abuse) don’t map cleanly, so the maintainers swapped in locally relevant examples and adjusted for Chinese rhetorical habits.

Key highlights

• Covers 24 AI markers across four categories: content, language/grammar, style, and communication/filler
• Includes a “slop vocabulary” watchlist: 此外, 深入探讨, 格局, 织锦, and other words that reek of LLM
• Explicitly not for evading AI detectors; the README insists the goal is better writing, not deception
• Adapted for Chinese context with rewritten examples (marketing copy, academic abstracts, blog posts)
• Runs entirely inside Claude Code as a skill; no external API or model calls

Caveats

• The project is essentially a translated rulebook plus prompt scaffolding; the heavy lifting is Claude’s, not custom code
• Only two files (SKILL.md, README.md), so the “installation” is just dropping them into ~/.claude/skills/
• No automated validation that the rewrite actually reduced AI markers; you eyeball it

Verdict

Worth installing if you write or edit Chinese text with Claude and want a systematic checklist for de-slopping it. Skip it if you expected a standalone model or automated scoring—you’re buying a style guide that talks back.