Skip to content

Humanizer 2.0

StoryScope Discourse & Anti-Slop Engineering Pipeline

Context
AI Agent Skill & Python CLI
Type
AI Product
Year
2026
Timeline
30 NarraBench Axes · 4-Phase De-Slop Pipeline
Artwork for Humanizer 2.0

In plain language

Readers can spot generic AI writing instantly—not because of particular vocabulary, but because AI constantly lectures, sums up with preachy life lessons, and uses robotic formulas. Humanizer 2.0 strips out the corporate jargon and artificial summaries, leaving behind crisp, direct, and credible writing that commands respect from clients and industry peers.

What it means for a business Protects company reputation and editorial credibility by eliminating the synthetic, generic AI tone that alienates sophisticated clients and readers.

Measured results

30
NarraBench Discourse Axes Audited
4
Sequential De-Slop Rewrite Phases
0%
Preachy Thematic Wrap-Up Sentences
2,400
Reference Corpus Texts Calibrated
Swapping banned words like "delve" or "tapestry" does nothing when the underlying paragraph structure still screams synthetic AI moralizing.

First-generation AI text rewriters rely on simplistic word blacklists. Yet empirical discourse research proves that readers recognize artificial text through macro-structural habits rather than individual words: thematic over-determination (wrapping every section in a neat, preachy moral), formulaic emotional tropes and melodrama, and rigid linear cause-and-effect transitions.

A four-phase discourse engineering pipeline and Python diagnostic tool based on the COLM 2026 StoryScope benchmark (30 NarraBench features across 2,400 texts) that strips synthetic AI writing habits, moralizing wrap-up paragraphs, and robotic phrasing from technical publications.

I translated the COLM 2026 StoryScope benchmark findings into an automated four-phase rewrite pipeline and companion Python linter. Phase 1 removes narrator thesis announcements and preachy summary conclusions (Story-Thema-Clar & Story-Ending). Phase 2 replaces synthetic visceral clichés with concrete behavioral facts. Phase 3 breaks formulaic linear chains, and Phase 4 grounds abstract assertions in specific named tools, release dates, and primary source citations.

  • Automated pattern detection for thematic over-determination (Story-Thema-Clar) and moralizing closing paragraphs
  • Somatic cliché scanner identifying synthetic visceral tropes and replacing them with objective operational state
  • Named-entity density auditor verifying that technical claims cite concrete tools, versions, and verifiable numbers

Implemented across engineering documentation, case studies, and editorial essays, Humanizer 2.0 eliminates artificial summary wrap-ups and restores authentic, authoritative practitioner voice verified against all 30 NarraBench discourse features.

How it was built, step by step.

How we translated COLM 2026 StoryScope discourse research into a 4-phase structural anti-slop engine.
  1. Stage 01 · Discourse Taxonomy

    Calibrating 30 NarraBench Axes Across 2,400 Corpus Texts

    Mapped the structural signatures of synthetic LLM prose beyond word blacklists: thematic over-determination (Story-Thema-Clar), moralizing tail wrap-ups, and autonomic somatic clichés.

    TAXONOMY LOCKED30 Discourse Axes

  2. Stage 02 · 4-Phase Rewrite Engine

    Sequential Structural De-Slop & Entity Grounding Pipeline

    Engineered the 4-pass Python diagnostic and rewrite harness: stripping preachy thesis bows, purging visceral tropes, breaking rigid "because -> therefore" chains, and anchoring claims in named primary sources.

    PIPELINE GREEN0% Moralizing Tails

  3. Stage 03 · Agent Skill & CLI UX

    Cross-Agent Skill Packaging & Side-by-Side Discourse Diffs

    Packaged the engine as an on-demand Claude Code, Antigravity, and Gemini CLI skill with side-by-side before/after structural scorecards.

    INTERFACE VERIFIED2,400 Texts Calibrated

  4. Stage 04 · Blind Evaluation & Ship

    Blind Editorial Verification Across Technical Essays & Docs

    Audited output across technical essays, documentation, and case studies, verifying zero synthetic wrap-up paragraphs and 100% citation preservation.

    SHIPPED TO PROD100% Fact Integrity

Stack

  • Python
  • NLP
  • Discourse Engineering
  • Claude & Gemini Skill
  • Editorial
  • Print

Credits

Research Synthesis & Engineering — Norbert Brett (br3tt) · Foundation — StoryScope (COLM 2026) / NarraBench · Stack — Python, NLP & Agent Skill Spec