write-humanlisted

# Write Human You take a draft and rewrite it so it reads like a specific person actually wrote it. Two things happen at once: the writing stops tripping AI detectors, and it gets better to read. Those are the same job, which is the whole insight behind this skill. ## The one idea that makes this work Most "humanizers" hand you a banned-word list and tell you to swap "delve" for "explore." That helps a little and misses the point. Word-swapping leaves the text flat, evenly paced, and opinion-free, which still flags as AI and now also reads worse. Here is what detectors actually measure, and it is not "machine-ness." Recent work (see `reference/why-detectors-flag-text.md`) found that detectors classify the raw output of *base* language models as overwhelmingly human, while heavily penalizing *instruction-tuned* output. In plain terms: detectors are tracking the residue of RLHF training, the habits a model picks up from being made polite, balanced, evenly structured, and careful. They score two things above all: - **Perplexity**: how predictable each next word is. AI takes the safest, most likely path every time, so the score stays unnaturally low. Humans surprise the reader. - **Burstiness**: how much sentence length and structure vary across a passage. AI clusters tightly around an average. Humans swing from three-word fragments to long, winding clauses. So the move is not "remove forbidden words." The move is: **rewrite so the text reads like a specific person continui