humanizer-pro (bridge)

This skill does not reimplement humanization. It points at the Better Humanizer — a measured, self-improving rewriter that is maintained as its own project so it never has to be kept in sync in two places.

• Upstream (single source of truth): https://github.com/jodyjdc/better-humanizer
• In this repo: the external/better-humanizer git submodule (version-pinned). Initialize it once with git submodule update --init --recursive.

When install.sh / install.ps1 runs, it installs the full Better Humanizer on top of this bridge, so in Claude Code /humanizer-pro is the complete tool — scorer, corpora, judges, registers, and personas — not this pointer.

Why it replaced the old humanizer

The previous humanizer skill was a static checklist of ~33 known AI "tells" from Wikipedia's Signs of AI writing. Better Humanizer keeps that pattern knowledge and adds the three things a checklist cannot do:

1. Measurement, not vibes. scripts/stylo.py (standard-library, no deps) rates every rewrite against a real human reference band: sentence-length burstiness, lexical diversity (MTLD), a function-word fingerprint, punctuation rates, tell counts, and document-level discourse structure.
2. Register awareness. "Human" is register-specific. Seven calibrated registers ship — for a novel studio the one that matters is literary (it tolerates ~13× the em dashes of casual prose, expects high burstiness, and so on).
3. Anti-over-correction. Every band has a floor and a ceiling. Scrubbing a chapter to zero em dashes, flat rhythm, or zero contractions is itself a tell, and it is penalized as a self_tell.

How to use it in the book pipeline

Run it on a chapter (or a passage) after drafting, before scoring — the same slot the old humanizer occupied in Phase 3 polish and the Phase 4/5 editorial passes.

Claude Code

/humanizer-pro
# then give it the chapter text (or path) and register: literary

Codex / any file-aware agent — open external/better-humanizer/ and follow SKILL.md (it routes through AGENTS.md). The loop is platform-agnostic; the scoring is a deterministic Python subprocess, so the result is identical.

Measured human-band gate (this repo's addition). The runner wraps the scorer so you can gate a whole manuscript on human-likeness, alongside the Genesis Score:

ubss humanize-score <project>            # scores manuscript/chapters, writes a report
ubss humanize-score <project> --gate     # exit non-zero if any chapter is flagged
ubss humanize-score chapter-03.md --register literary --json

A flag is a measured signal to inspect, not an automatic reject: re-run the Better Humanizer loop on the flagged chapter, then re-score. The distance should drop and the self-tells clear. See docs/humanizer-pro-integration.md for the full contract and the honest calibration caveat.

Rule

Do not paste humanization logic, lexicons, corpora, or scoring code into this repo. If the technique needs to change, change it upstream in jodyjdc/better-humanizer and bump the submodule pointer here (git submodule update --remote).