ai-humanizer

A Claude Code skill that lowers the AI-detection score of English text without changing what it says — every key term, number, and citation is preserved verbatim, and the original logic stays intact.

It works by detector-guided surgical paraphrase (cf. arXiv:2506.07001): a detector flags which sentences read as AI, the agent rewrites only those, re-checks, and loops — keeping the best result and never dropping a protected term.

• No API key. The baseline detector is the keyless ZeroGPT endpoint; the agent running the skill is the rewriter, so there is no separate paraphrase API to configure.
• Term-safe. Numbers, percentages, citations, drug/gene names, theorems, standards, and inline notation are extracted and verified character-for-character after every round.
• Honest. Real human academic prose scores 0–50% on these detectors, not 0. Saturated topics (clinical, monetary policy, common legal/historical essays) are reported truthfully instead of over-promised. ZeroGPT is treated as a free proxy — Turnitin is the ground truth.
• Reads detector reports. Hand it a Turnitin / GPTZero / Originality.ai PDF and it uses the report's highlighted spans as the attack list.

⚠️ For legitimate use — reducing false positives on your own writing, de-roboticizing AI-assisted drafts, and detector research. Follow your institution's and publisher's policies on AI assistance.

Install

/plugin marketplace add recomby-ai/ai-humanizer
/plugin install ai-humanizer@ai-humanizer

Update later with:

/plugin marketplace update ai-humanizer

Manual install (other agents — Codex, Cursor, Gemini CLI)

The skill is plain Node with no dependencies. Clone the repo (or copy the skill folder) and point your agent at SKILL.md:

git clone https://github.com/recomby-ai/ai-humanizer
# the skill lives at:
#   ai-humanizer/plugins/ai-humanizer/skills/ai-humanizer/

Then tell the agent to read SKILL.md and follow it. The two scripts run under any Node 18+:

node scripts/detect.mjs <file>            # AI score + flagged sentences (JSON)
node scripts/terms.mjs extract <file>     # protected factual terms
node scripts/terms.mjs verify <orig> <new> <terms.json>   # term-preservation gate

How to use it

Trigger phrases (English or Chinese): "humanize this", "lower the AI score", "降AI率", "降低Turnitin AI率", "make this read human", "reduce GPTZero score" — or just hand over a draft, optionally with a detector report PDF.

The skill will:

1. Detect a baseline score and the flagged sentences.
2. Build the protected-term list (deterministic regex + the agent's own domain-entity pass).
3. Rewrite only the flagged sentences under a set of surgical rules grounded in measured findings (see reference/principles.md).
4. Re-detect and loop (up to ~5 rounds), keeping the best version.
5. Verify every protected term survived, then show a before → after report with a sentence-level diff.

Examples

Real before/after runs are in examples/:

• Nursing essay — normal academic domain, 100% → 11.1% in one round, 7/7 terms kept.
• Clinical meta-analysis — a detector-saturated domain. Two independent rewrites scored 0% and 26.5% under identical rules (14/14 terms kept) — the clearest illustration of why best-of-N matters and why the skill won't over-promise.

Why these rules (the short version)

The detailed evidence is in reference/principles.md. The headline findings, measured against ZeroGPT with confirmed-human baselines:

• The signal is lexical predictability, not connectives. "Furthermore / Moreover / plays a crucial role" is a false tell — a Band-9 IELTS sample full of them scored 19.6%; a polished AI clinical review with none scored 100%.
• Fluency is the tell. Polished, frictionless prose reads as AI. You lower the score by allowing slight roughness and longer, idea-dense sentences — not by chopping text into short punchy lines.
• For academic content, more precision — not more casual. Swap high-probability defaults for precise, lower-frequency domain vocabulary. "Explain it to a friend" is the wrong move here.
• Best-of-N is the biggest lever. Score variance is sampling-dominated (the same input can swing 0–52%). Generate a few independent rewrites and keep the lowest scorer.

License

MIT © recomby-ai