ai-humanizerlisted

# ai-humanizer Reduce the AI-detection score of English text **without** changing what it says. This is detector-guided surgical paraphrase: a detector tells you *which* sentences read as AI, you rewrite only those, you re-check, you loop. Everything not flagged stays byte-for-byte. Every protected term survives verbatim. This skill is self-contained Node (no API key, no install): the detector is the keyless ZeroGPT endpoint, and **the agent running this skill is the rewriter** — you replace what used to be a separate paraphrase API. ## The one idea that makes this work AI detectors do **not** key on "Furthermore / Moreover / plays a crucial role". That folklore is wrong and was disproven on real data (a Band-9 IELTS sample full of "Firstly/Furthermore/Consequently" scored 19.6%; a polished clinical review with none of them scored 100%). The real signal is **lexical predictability** — how default, high-probability, and *smooth* each next word is. **Fluency is the tell.** Polished, frictionless prose is what reads as AI. So you lower the score by making flagged sentences **less predictable and less smooth** while keeping them correct and meaning-identical: - Replace the highest-probability wording with **precise, lower-frequency** alternatives an expert in the field would actually reach for. - Prefer **longer, idea-dense sentences** with stacked subordinate clauses. Do **not** chop everything into short punchy sentences — that is its own AI tell. - Allow **slight rou