Skip to content

README.md

Latest commit

 

History

History
135 lines (100 loc) · 5.29 KB

README.md

File metadata and controls

135 lines (100 loc) · 5.29 KB
Raw

my-humanizer-es-en

A Claude skill that rewrites AI-flagged academic and professional text (Spanish and English) so it reads like a real person wrote it — and passes stylometric AI detectors like Compilatio Magister+, Turnitin, GPTZero and Originality.ai — while keeping the original meaning, data, and citations untouched.

Built from a real case: it took an academic article from 50% → 13% AI on Compilatio across three iterations, without changing a single citation.

What it does

Detectors flag rhythm and structure, not vocabulary. This skill breaks the uniform cadence of AI writing and adds human "burstiness":

  • Kills the statement: elaboration colon habit (the strongest tell).
  • Breaks symmetric enumerations ("two things:", "three fronts:").
  • Mixes very short sentences with long ones (varied length = human).
  • Adds first person and concrete, slightly messy detail where it fits.
  • Never touches citations, numbers, dates, or the reference list.

Two ways to use it

With a detector report (best). Give Claude the detector's PDF report and it reads the colored highlights (blue = AI, green = similarity) straight from the PDF, so it rewrites only the flagged passages, then rebuilds your .docx without breaking its formatting.

Without a report (blind mode). No report? No problem. The skill runs its own internal leak scan (scan_leaks.py) that scores every sentence for likely AI tells — long sentences, colon-lists, symmetric enumerations, AI vocabulary, em-dashes — ranks the worst ones, and rewrites those, then checks burstiness metrics. It's a heuristic proxy for a detector: it finds what probably still reads as AI so you can fix it before submitting anywhere.

Either way it keeps meaning, data and citations intact, and leaves the reference list alone.

Contents

my-humanizer-es-en/
├── SKILL.md                    # the skill (methodology + rulebook Claude follows)
└── scripts/
    ├── analyze_report.py       # report mode: extract blue (AI) / green (similarity) passages from a detector PDF
    ├── scan_leaks.py           # blind mode: heuristic AI-leak scanner (ES+EN), ranks the riskiest sentences
    ├── replace_paragraphs.py   # rewrite paragraphs inside a .docx, preserving formatting
    └── text_metrics.py         # measure burstiness: colon count, mean/stdev sentence length

What knowledge is inside

The skill is self-contained. It folds in everything used to take the reference article from 50% to 13%: the full Wikipedia "Signs of AI writing" pattern catalogue (Spanish + English), the empirical findings from a real Compilatio Magister+ loop (colon-lists and symmetric enumerations are the highest-value fixes), burstiness metrics, the blind-mode leak scanner, and format-preserving .docx editing. No other skills need to be installed for it to work.

Install

Option A — with the skills CLI (recommended)

npx skills add ljmor/my-humanizer-es-en

This installs it into ~/.claude/skills/my-humanizer-es-en/, available to both Claude Code and Claude Desktop.

Option B — manual

git clone https://github.com/ljmor/my-humanizer-es-en.git
cp -r my-humanizer-es-en ~/.claude/skills/my-humanizer-es-en

Python dependencies (used by the scripts)

python3 -m pip install pymupdf pypdf defusedxml

Updating (if you already have a previous version installed)

With the skills CLI:

npx skills update my-humanizer-es-en
# or, to force a clean reinstall of the latest from GitHub:
npx skills remove my-humanizer-es-en
npx skills add ljmor/my-humanizer-es-en

Manual: overwrite the installed copy with the latest from the repo.

git -C my-humanizer-es-en pull        # in your local clone
rm -rf ~/.claude/skills/my-humanizer-es-en
cp -r my-humanizer-es-en ~/.claude/skills/my-humanizer-es-en

Restart Claude Code / Claude Desktop (or reload skills) so the updated SKILL.md is picked up. Your previous version is simply replaced — nothing else to clean up.

Use

In Claude Code or Claude Desktop, just ask:

  • "humanize this for Compilatio"
  • "baja el porcentaje de IA de este artículo"
  • "parafrasea las secciones en azul de este informe"

…or run the skill directly with /my-humanizer-es-en. Hand Claude the document (.docx) and, if you have it, the detector's report PDF. Claude analyzes the report, rewrites the flagged text, rebuilds the document, and iterates with you until the score hits your target.

Honest limits (please read)

  • The score is an estimate, not a promise. AI detection is statistical. The skill may not drop the percentage as much as you hope on a given run, and no tool can guarantee a specific number. Expect to iterate.
  • The blind-mode scanner is a proxy, not a detector. It flags what probably reads as AI. Passing it is a good sign, not a certificate.
  • Institutional detectors are run by you. Compilatio/Turnitin inside an LMS can't be called by the skill. The workflow is a loop: it delivers a version, you submit it and share the new report, it targets whatever is still flagged — each round usually lowers the score further.
  • It never fakes citations or data. If the honest, human-sounding rewrite still trips a detector, the answer is another pass — not inventing sources.

License

MIT