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

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.

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

Vocabulary, punctuation, and phrasing. Inflated symbolism, promotional language, superficial "-ing" analyses, vague attributions, em dash overuse, rule of three, AI vocabulary, negative parallelism. Built from Wikipedia's Signs of AI writing.

The copy-tells.md reference adds the tells that show up specifically in public copy, ranked by a 3.2M-post Reddit analysis of what people actually flag.

Six audits run one at a time, because aspect-based checking found 95% of issues in the study's own pipeline against 68% for a single combined pass:

1. Theme explicitness. AI states its lesson. Narrator explains the theme 77% of the time against 52% for humans.
2. Structural tidiness. Single-track, everything resolved. Humans digress and leave threads open.
3. Emotion mode. The largest gap in the study. AI performs emotion through the body 81% of the time against 38% for humans. Humans just name the feeling.
4. Reference specificity. Humans name real things (47% against 24%). AI stays at vague allusion.
5. Reader engagement. Humans acknowledge the reader. AI writes as though no one is watching.
6. Shape convergence. Does this piece have the same skeleton as your last three?

Do not trade one default for another. The study's deepest finding is convergence: all five models occupy one tight region of structural space while humans are dispersed. Rarity is the human signal.

If every piece now opens mid-scene, names three feelings, and ends unresolved, you have built a new detectable cluster. Pick one or two interventions per piece, vary them across pieces, and be able to say why this piece got this shape.

git clone https://github.com/NulightJens/humanizer-stack.git
cd humanizer-stack
./install.sh

This symlinks both skills into ~/.claude/skills/, so updates land with a git pull. Pass --copy if you would rather have independent copies than symlinks.

To install manually, copy skills/humanizer and skills/structural-humanizer into ~/.claude/skills/ (user-level) or .claude/skills/ (project-level).

In Claude Code, the skills trigger on intent:

humanize this post
de-slop this lesson
run the structural pass on draft.md

Run the surface pass first, then the structural pass. docs/PIPELINE.md covers the order and what each layer owns.

Both scanners are deterministic and hook-friendly. They catch the pattern-matchable slice only, and neither replaces the judgment work in the skills.

python3 scripts/copy_scan.py draft.md
python3 skills/structural-humanizer/scripts/structural_scan.py draft.md

python3 scripts/copy_scan.py --json draft.md      # machine-readable
python3 scripts/copy_scan.py --strict draft.md    # exit 1 on any hit
cat draft.md | python3 scripts/copy_scan.py -     # stdin

Mark a line copy-ignore to suppress an intentional usage.

• StoryScope studied roughly 5,000-word fiction. Applying it to short nonfiction is an inference, not a result the paper establishes. The subset that transfers most cleanly is audits 1, 3, 4, and 6.
• Nothing here makes text undetectable, and that is not the goal. The goal is writing that reads as though a person with a specific point of view wrote it, because a person did.
• The scanners catch maybe half. Cadence, formulaic shape, and polished-but-empty filler are only visible to a human reader.
• Audit 3 contradicts standard writing advice. "Show, don't tell" is now a machine signature. That is what the data says, and it is worth sitting with before applying.

Built on work by @blader (MIT), jcarterjohnson (MIT), and Wikipedia's WikiProject AI Cleanup (CC BY-SA 4.0). Grounded in Russell et al. 2026.

Full breakdown with license obligations: ATTRIBUTION.md.

MIT for this repository's own work. Portions carry upstream terms, including CC BY-SA 4.0 material with share-alike obligations. See ATTRIBUTION.md before redistributing.