moai-domain-humanize

Run a humanization pass that detects catalogued AI tells — formulaic connectors, cliché openers, inflated significance — and rewrites only the flagged spans. This Skill supports Korean, English, Japanese, and Chinese with per-language tell catalogues and severity-based removal thresholds.

Common tells include mechanical transition stacking, rule-of-three tricolons, vague attribution like "experts argue", forced uplifting endings, and formatting abuse such as excessive bold. Each tell carries a severity tier: S1 tells are removed on first occurrence, S2 at three or more instances.

Yes, the Skill ships dedicated modules for Korean, English, Japanese, and Chinese, each with language-native prose and copy-layer catalogues. Copy tells do not transfer mechanically between languages, so each module applies its own rules and examples.

No — meaning preservation is the top rule. An Invariant Ledger records facts, numbers, names, and caveats before editing, and a post-edit Delta Audit rolls back any edit that alters a supplied item. Any meaning-distortion flag forces the lowest quality grade.

Automated detectors are unreliable across these four languages, especially on CJK polite registers where they false-positive on correct formal writing. This Skill is a pattern-based editing tool that relies on catalogued tell categories and clustering, not on a detector's verdict.

Use the copy-review QA gate when copy came from another tool or workflow and must be reviewed before application. It detects formulaic copy, proposes at least three alternatives per finding, and never auto-applies changes — accepted fixes then flow through the standard rewrite mode.