Humanizer Pro — AI skill for Claude Code

humanizer-pro

Agent skill. 44 constraints for writing text that doesn't read like AI output. Composition-time rules first, with a preservation mode for editing drafts you didn't write. Works with any agentic framework that supports SKILL.md.

What it does

Most "humanizer" tools work backward: you write with an LLM, then run the output through a filter that swaps words and restructures sentences. The result still reads like processed AI text, just with different vocabulary.

This skill works forward. It loads 44 constraints before writing starts and applies them per-sentence during composition. The output is written human from the first draft, not cleaned up after the fact.

When the input is an existing draft rather than a writing task, preservation mode inverts the priority: the constraints still apply, but the author's voice outranks them.

How it works

Five phases run in sequence:

1. Calibration locks the voice profile (register, perspective, stance, audience) before any writing begins. Optionally calibrates from a writing sample to mirror the user's existing voice.
2. Composition applies all 44 constraints as active rules. Each sentence is shaped by them during generation, not checked against them afterward.
3. Voice injection adds the things constraint-following alone can't produce: sentence rhythm variation, opinion insertion, concrete specifics, deliberate imperfection.
4. Verification runs four passes: a grep-based pattern scan for known AI tells (graded HIGH/MED/LOW by density and position, not mere presence, with hits inside quotes and code discarded first), a structural audit (sentence length variance, paragraph openers, confidence variation, noun-verb ratio, plus a substance spot-check that runs deletion and inversion micro-tests on the weakest paragraphs), an introspective self-audit ("what makes this still sound AI?"), and a read-aloud test. A calibration rule keeps editing proportional: thin clusters rather than scrubbing e