blader/humanizer
A Claude skill that finds the habits AI leaves in English prose, like vague sourcing, puffed-up importance and stock phrases, and rewrites them. It keeps the facts as they were, matches a sample of your own writing if you give one, and leaves code, tables and links alone. The idea of a fixed checklist of AI tells, plus a second pass that checks the rewrite against the source, is worth borrowing.
Idea adaptedTested by others
I learnt this repo's idea and adapted it into my own workspace setup.
How we know
Someone outside the project tested it. Their result is below, with a link to the original post. We have not tested it ourselves.
Good for
Writers who draft in English and want a checklist of AI tells to scrub out before posting.
Not for
Vietnamese text. The patterns are written for English, so they may miss the tells in Vietnamese.
Sources
Desklib detector catch rate fell from 62.7% to 32.5%, Fast-DetectGPT 63.5% to 47.6%, MAGE 45.2% to 23.8%; still flagged: 47.6% on three detectors
Tool Benchmarks · · how: Claude Opus 5.5 texts rewritten with the skill, then scored by 7 AI-text detectors before and after; number of texts not stated on the page
https://toolbenchmarks.com/humanized-ai-text/85% of 179 unsupported claims in the drafts survived the skill (88% with a plain one-line prompt); 0.86 AI tells per 100 words vs 1.26 for the plain prompt; about 5.8x the cost per call
Hard Numbers (Pitambar Mahato) · · how: 24 chatbot-style drafts through Claude Code (Opus 5.5) with skill v3.1.0 vs a one-line control prompt, Sonnet 5.5 judge; one run per cell; unsupported claims were counted by an LLM, and 85% vs 88% is not a real gap
https://hardnumbers.dev/articles/blader-humanizer-benchmark-it-keeps-what-drafts-invent8.5% average AI-detector score after the rewrite; 9 of 10 posts under 20%
Figuring Out AI (Ankit Vora) · · how: 10 AI-written blog posts rewritten with the skill in Claude Code, scored with ZeroGPT (named for the first post; the page does not name the detector for the rest); no control arm; an early version of the skill
https://ankitvora2.substack.com/p/day-1-of-testing-ai-agents-humanizer
Repo
- Stars
- 40,372
- License
- MITUse it, change it, sell it, closed-source included. Keep the author's copyright notice.
- Maturity
- active
- Last push
- 2026-08-19
- Last checked
- 2026-11-14