// the find
blader/humanizer
Agent skill that removes signs of AI-generated writing from text
A Markdown-based agent skill that rewrites AI-generated text to sound more human, based on 33 patterns derived from Wikipedia's AI writing detection guide. It works in Claude Code, Cursor, Codex, and any harness that supports skill-style instructions. The whole runtime artifact is a single SKILL.md file.
The 33-pattern taxonomy is genuinely well-researched and specific — naming concrete tells like em dash overuse, synonym cycling, and 'rule of three' is more useful than generic 'sound less robotic' advice. The no-fabrication rule is a good constraint that prevents the common failure mode of LLMs inventing plausible-sounding details during rewrites. Voice calibration from a writing sample is a practical feature that separates this from most prompt-based rewrite tools. The before/after examples in the README are honest and illustrative — the Lisbon example actually demonstrates what humanized writing looks like rather than just describing it.
At 30k+ stars this is clearly viral, but the core artifact is a Markdown prompt file, not code — there's nothing here that couldn't be copy-pasted into a system prompt, and the GitHub repo is mostly packaging and distribution tooling around that fact. The 'skills CLI' ecosystem (skills.sh, npx skills add) is a third-party dependency with unclear longevity; if it disappears, the install story breaks. The pattern list is derived from Wikipedia's editorial guidance, which was written for encyclopedia articles, not general prose — several patterns (title case headings, citation vagueness, notability name-dropping) are specific to that context and less relevant for code documentation, blog posts, or emails. There's no evaluation harness or benchmark showing whether this actually fools AI detectors or improves human ratings, so effectiveness is anecdotal.