A Claude skill that rewrites AI-generated text so it reads as natural human writing and holds up against AI detectors. It goes beyond surface "signs of AI writing" and targets the signals detectors actually measure, then ships a scorer that estimates a before/after detection-risk number.
Honest disclaimer. AI detectors are unreliable and uncalibrated. A Stanford study (Liang et al., 2023) found they flagged 61% of non-native-English essays as AI. There is a proven impossibility result: as models approach human text, the best detector approaches a coin flip. This tool is for legitimate writing improvement — polishing AI-assisted drafts and helping people who are falsely flagged. No method beats every detector, and trained classifiers (Pangram, GPTZero 3.2b) may be unbeatable by tooling. Use it honestly.
# install for your user (~/.claude/skills)
npx ai-humanizer-skill install
# or into the current project (./.claude/skills)
npx ai-humanizer-skill install --projectThen restart Claude Code and run /ai-humanizer, or just ask Claude to "humanize this text".
Other commands:
npx ai-humanizer-skill where # show install destination
npx ai-humanizer-skill uninstall # remove it
npx ai-humanizer-skill --helpnpx installs into Claude Code only. To use the skill in the Claude.ai web app, upload it as a ZIP:
- Download
ai-humanizer.zipfrom the latest release. - In Claude.ai, go to Settings → Customize → Skills (also labeled "Capabilities" on some plans).
- Click + Create skill and upload
ai-humanizer.zip. - Toggle the skill on, then ask Claude to humanize text.
Requirements: custom skill uploads need a Pro / Max / Team / Enterprise plan with code execution enabled (Settings → Capabilities). Code execution also lets the bundled score.py run in the web sandbox, so before/after scoring works there too. Uploaded skills are private to your account.
AI detectors split into two families that need opposite tactics. The skill handles both:
| Family | Examples | Beaten by |
|---|---|---|
| Statistical / zero-shot | DetectGPT, Binoculars, GPTZero's perplexity component, most free tools | Genuinely raising perplexity (specific diction) and burstiness (varied sentence length), plus a real structural rebuild |
| Trained neural classifier | Pangram, GPTZero 3.2b, Originality.ai, Turnitin, Copyleaks | Matching a real human voice + a detector-feedback loop. Synonym swaps and generic humanizers are adversarially trained into these. |
Six levers: raise perplexity, raise burstiness, rebuild logic (not words), match a real voice, strip Unicode/formatting watermarks, and recursive/back-translation paraphrase for watermark removal. It refuses to break facts or over-edit formal writing.
The skill bundles scripts/score.py (Python 3, standard library only). It estimates a 0–100 detection-risk score from measurable signals — burstiness (σ/μ of sentence length), lexical-tell density, Unicode artifacts, and vocabulary richness.
python ~/.claude/skills/ai-humanizer/scripts/score.py mytext.txt
echo "some text" | python ~/.claude/skills/ai-humanizer/scripts/score.pyExample movement from the included test cases:
| Case | Risk before | Risk after |
|---|---|---|
| Free prose | 75/100 (likely AI) | 0/100 (likely human) |
| Technical | 67/100 | 4.6/100 |
| Math proof | 47/100 | 7.3/100 |
Caveat: the score estimates the statistical detector family only. A low score does not guarantee a trained classifier (Pangram, GPTZero) will pass the text. The only ground truth is running a real detector and feeding its sentence-level highlights back in — the skill's detector-feedback workflow does exactly that.
skill/
├── SKILL.md operational guide (the skill Claude loads)
├── scripts/score.py before/after detection-risk estimator
└── reference/
├── detector-landscape.md per-detector teardowns + research citations
└── test-cases.md three worked before/after examples with scores
- Node.js >= 16 (for the installer)
- Python 3 (only for the optional
score.py) - Claude Code, or any host that loads
~/.claude/skills
MIT © mlswijerathne