Human-AI

Overview

Strips mechanical writing signals from English text and restores rhythm, precision, and voice. Combines pattern detection (43 patterns across 3 tiers), statistical rhythm measurement (burstiness, TTR, entropy), and voice injection into a single iterative skill with scoring.

The goal is a better text, not a fooled detector: no rewrite can guarantee that a tool will classify the result as human, and AI-detector scores are not a quality criterion — those tools misfire often and penalize neurodivergent and non-native writers disproportionately.

The statistical metrics serve as measurable proxies for natural rhythm, not as a scoreboard to beat. The research behind the skill is useful for a different reason: it shows that swapping synonyms changes nothing, because the problem lives in the sentence’s architecture — length, rhythm, clause structure, and information density.

Companion to the humanizar skill (PT-BR). For Portuguese text, use humanizar.

When to Use

• English text sounds generic, bland, or AI-generated
• Requests like “humanize”, “de-slop”, “remove AI patterns”, “make it sound human”
• Rewriting with voice and personality, or fixing a generic tone
• Agent pipelines generating English content
• Review another agent’s output before publishing

What Makes It Different

Feature	blader/humanizer	brandonwise/humanizer	human-ai
Patterns	29	28 + statistical	43 (incl. P31-P43 emerging 2026)
Iterative scoring	❌	❌	✅ 0-100 with loop
Metrics script	❌	❌	✅ scripts/measure.py
Voice presets	❌	❌	7 presets
Anti-synonym-swap	❌	❌	✅ Structural rewrite enforced
Empirical baselines	❌	Partial	✅ SSRN + GPTZero + NeurIPS

Operation Modes

Mode	When to use	Result
full_mode	Default — “humanize this”	Metrics + diagnosis + rewrite + scoring
direct_mode	Pipelines or “humanize quick”	Final version + synthetic report
review_mode	Audit text from another agent	Aggressive rewrite + before/after metrics

Installation

npx skills add https://github.com/fabricioctelles/skills --skill human-ai

Research Foundation

Source	Finding
RAID Benchmark (ACL 2024)	Structural paraphrasing drops DetectGPT from 70.3% → 4.6%
humanizerai.com (2026)	Vocabulary bans HURT performance by 43 percentage points
GPTZero	Burstiness (sentence length variance) is the primary detection signal
SSRN	Human TTR: 0.553 vs AI: 0.455
NeurIPS 2023	Human intrinsic dimensionality ~9 vs AI ~7.5
Washington Post	“It’s not X, it’s Y” = #1 AI tell across 328K messages

Part of this research was produced by measuring detector bypass rates. The skill uses those findings for what they reveal about rhythm and structure — not as a target. None of these numbers proves human authorship, and none of them decides what gets rewritten.

Changelog

v1.0.1 (Aug 2026)

• Positioning aligned with humanizar: the skill no longer promises “undetectable” text, and AI-detector scores are no longer a criterion
• Statistical metrics repositioned as proxies for natural rhythm, not a scoreboard
• Dropped the contraindication based on “text already validated as human by multiple detectors”

v1.0 (Jul 2026)

• Initial release with 43 patterns (P1-P43), including 13 emerging from 2026
• 3 operation modes with iterative scoring (0-100) and strategy fallback
• 7 voice presets (Essay, Journalistic, Academic, Corporate, Social, Casual, Legal, Instructional)
• scripts/measure.py for deterministic metric calculation (zero dependencies)
• 7 documented gotchas from real operational failures
• Empirical baselines calibrated from published research
• Bidirectional cross-reference with humanizar skill (PT-BR)
• Based on blader/humanizer, brandonwise/humanizer, and Aboudjem/humanizer-skill

📄 Full documentation on GitHub

---

TypeSafe Jev Integration

This skill supports optional integration with TypeSafe Jev — a judgment engine providing calibrated probability assessments for subjective criteria.

When to Use Jev

Evaluation	Use Jev?	Method
Pattern detection	No	Deterministic (regex, patterns)
Pattern severity	Yes	Score (0.0-1.0)
Voice match quality	Yes	Score
Rewrite confidence	Yes	Score
Change classification	Yes	Noul (structural/cosmetic/hybrid)

Available Questions

• 8 Score questions: pattern severity, voice adequacy, fluency, tone preservation, naturalness, restored entropy, achieved burstiness, overall confidence
• 12 Noul questions: change type, pattern category, intervention level, factual preservation

Discovery Protocol

from typesafe import jev_available

if jev_available():
    from typesafe import Score, Noul
    # Use Jev for subjective criteria
else:
    # Fall back to heuristic scoring

When Jev is unavailable, the skill uses internal heuristic scoring — functional but less nuanced.

📄 See jev_questions.json | See jev-integration.md