humanizerlisted

# Humanizer v3.0.0 Two engines in one skill: 1. **Pattern engine.** Removes the ~60 patterns that mark text as AI-generated, while preserving meaning, register, and length. 2. **Integrity engine.** Enforces honest claims: no fabricated metrics, no unprovable superlatives, credible sourcing, self-consistent numbers. A piece ships only when both pass. Human-sounding lies and honest robot-prose are both failures. ## What this skill is and isn't The patterns flagged here are statistically more common in LLM output, but humans on autopilot produce the same shapes, especially under deadline pressure, in unfamiliar genres, or in a second language. Independent audits found false-positive rates above 60% on non-native English writers (Liang et al., Stanford, Patterns 2023) and misclassification above 70% on open-source detectors (Jabarian & Imas, BFI WP 2025-116). Adversarial paraphrase cuts detection accuracy by ~88% across every method tested (arXiv:2506.07001). So: signals, not proof. Act on **clusters**, never a single hit. Never use this skill as the sole basis for a consequential decision about a human writer (grades, hiring, attribution). ## Modes - **rewrite** (default): audit, then return a clean version. - **detect**: flag only, grouped by severity, with a judgment note per flag (clear problem vs. possibly intentional). Trigger on "detect", "flag only", "audit", "scan", "what AI patterns are in this". - **edit**: fix a named file in place with minimal targeted edits.