humanizer-prolisted

Measured, self-improving rewriter. Generates several candidate rewrites, scores each with a hybrid scorer (std-lib stylometrics + an LLM judge panel) against the real human distribution for a register, applies vetoes, keeps the best, and iterates on the judges' critique. Targets the human band (floor AND ceiling) so it never over-corrects into a machine-laundered tell. v1 register: spontaneous. Explicit non-goal: defeating commercial AI detectors.
jodyjdc/better-humanizer · ★ 0 · AI & Automation · score 68

Install: claude install-skill jodyjdc/better-humanizer

# humanizer-pro Rewrite AI-sounding text into prose that sits inside the human distribution for its register, and **prove** it with a score instead of trusting vibes. This is the measured successor to the original `humanizer` skill: same pattern knowledge (now in `lexicons/ai_tells.json`), plus measurement, register awareness, and anti-over-correction. **Platform note.** Claude Code loads this file as the `/humanizer-pro` skill. Codex users enter through [`AGENTS.md`](AGENTS.md), which maps the Claude Code tool names below (`Bash`, `Read`/`Write`/`Edit`, `AskUserQuestion`, `Workflow`) to their Codex equivalents. The loop itself is platform-agnostic and the scoring is a deterministic Python subprocess, so the result is identical on both. ## Inputs - `text`: the passage to humanize (the SOURCE). - `register`: default `spontaneous`. Selects `registers/<register>.md` (the brief) and `corpus/<register>/reference-stats.json` (the numeric bands). - optional **target overrides** (pass the matching flag to `scripts/stylo.py`): - `expertise` — `--expertise novice|practitioner|expert` (readability tier within the register; default practitioner = the full register). - `voice sample` — the user's own writing: run `scripts/build_reference.py --voice-sample <dir> --label <name>` first, then score with `--voice <name>` so the human target becomes *them*. - `persona` — `--persona <name>` (e.g. `reddit-power-user`, `seasoned-journalist`, `startup-founder`, `academic