0xwilliamortiz
Featured
openclaude-improved
OpenClaude is an open-source CLI coding agent that runs on any platform and supports a wide range of LLM providers, offering the same tools and workflows as Claude Code.
humanizer-stack
A two-pass pipeline (surface words + structural discourse) packaged as Claude Code Skills to remove signs of AI writing from outward-facing text, grounded in the StoryScope study showing structural tells are more durable than surface ones.
Stars
121
7-day growth
No data
Forks
8
Open issues
0
License
NOASSERTION
Last updated
2026-07-24
AI repository intelligence
FR-AI / ANALYSISWhy it is worth attention
It tackles the overlooked structural dimension of AI writing fingerprints, backed by peer-reviewed research (93.2% F1 detection on discourse features) and offers deterministic scanners plus a free community for practical tooling.
Who it is for
- Claude Code users seeking to humanize AI-generated text
- Content creators and marketers wanting natural-sounding copy
- Writers and editors in fiction or long-form nonfiction
- AI researchers interested in detection and humanization techniques
Use cases
- Humanizing AI-generated blog posts, articles, or marketing copy by running pass 1 then pass 2
- Using the deterministic scanners to quickly flag surface-level or structural AI tells in drafts
- Integrating the skills into a Claude Code workflow for iterative rewriting
- Learning about structural AI tells through the provided references and genre calibration
Strengths
- Two-pass approach (surface + structure) addressing the most durable AI fingerprints, supported by the StoryScope study
- Deterministic scanners (copy_scan.py, structural_scan.py) that are hook-friendly and machine-readable
- Open source (MIT) with clear attribution and a free community for support and discussion
- Includes detailed references (copy-tells, storyscope findings, genre calibration) for educated decision-making
Considerations
- The core research studied ~5,000-word fiction; applying to short nonfiction is an inference not directly validated
- Scanners catch only about half of AI tells; cadence, formulaic shape, and empty filler remain human-judgment zones
- Audit 3 ('Show, don't tell' is now a machine signature) contradicts standard writing advice and requires careful application
README quick start
humanizer-stack
A two-pass pipeline for removing the signs of AI writing from outward-facing text, packaged as Claude Code Skills.
Most humanizers only fix words. That is the easy half, and it is the half that is decaying fastest. This repo pairs a surface pass with a structural pass, because the research says structure is where the durable fingerprint lives.
Free community. I build tools like this in the open inside the Jens AI Community, a free Skool group for putting AI to work in your business. If this repo is useful to you, come join us: https://www.skool.com/jens-ai-community-1306
Why two passes
The StoryScope study (Russell et al., 2026) classified 61,608 stories from humans and five LLMs using only discourse-level features, with every style feature withheld. It detected AI text at 93.2% F1.
Then the authors ran AI text through LAMP, a professional span-level rewriting system that strips cliche, purple prose, and redundant exposition. Functionally, a very good surface humanizer.
Detection dropped 1.6 points.
Meanwhile the surface layer is eroding on its own. GPT 5.4 already cut its em-dash usage sharply, and fine-tuning drops stylistic detection from 97% to 3%. Word-level tells are a moving target. Structural tells require structural rewrites.
So: pass 1 fixes the words. Pass 2 fixes the shape. Run them in that order.
What is in here
skills/
humanizer/ Pass 1: words and phrasing
SKILL.md
references/copy-tells.md Copy-specific tells (em dash, hype vocab, antithesis)
structural-humanizer/ Pass 2: discourse structure
SKILL.md
references/
storyscope-findings.md The study distilled: 30 core features with rates
genre-calibration.md Which audits apply per genre
scripts/structural_scan.py Deterministic scanner for grep-able structural tells
scripts/
copy_scan.py Deterministic scanner for mechanical copy tells
docs/
PIPELINE.md How the passes chain, and what each one owns
Pass 1: humanizer
Vocabulary, punctuation, and phrasing. Inflated symbolism, promotional language, superficial "-ing" analyses, vague attributions, em dash overuse, rule of three, AI
Description
Two-pass pipeline for removing AI writing tells from outward-facing text: a surface pass plus a structural pass grounded in the StoryScope study. Packaged as Claude Code Skills. Free community: skool.com/jens-ai-community-1306
Related repositories
Similar projects matched by category, topics, and programming language.
S40911120
Featured
recensa
Recensa is a self-hosted web viewer that indexes Claude Code session transcripts into a local SQLite database, enabling full-text search, replay, and audit of all past agent conversations without uploading data anywhere.
makecindy
Featured
cindy
Cindy is an open-source AI agent that runs locally on your machine, integrates multiple AI harnesses and models, and provides memory, skills, and automation to perform real work in your projects and apps.