humanizer

Humanizer is an agent skill, distributed as plain Markdown, that rewrites text which sounds AI-generated so it reads as if a person wrote it, without altering the underlying meaning. Because it is only Markdown, it works with any agent that supports skills. Installation is done through the Skills CLI with `npx skills add blader/humanizer --global`; omitting `--global` installs it only in the current project, and `--agent <name>` or `--agent '*'` selects which agents receive it. Claude Code 2.1.142 or newer can install it as a plugin via the marketplace commands, and Claude Desktop users can upload the repository as a ZIP or copy `SKILL.md` into the agent's skill folder. The skill responds to `/humanizer` (or `/humanizer:humanizer` as a plugin). Usage is straightforward: call the skill and paste text, ask in plain language, or point it at a file path to rewrite prose in place. If a writing sample is supplied, the rewrite follows that sample's rhythm, word choice, punctuation, and deliberate quirks, including dashes. When given a file, it changes only prose and leaves code, data, frontmatter, and link targets untouched. The skill's rationale is that language models default to the most statistically likely continuation, which fits the widest range of readers and subjects, whereas a person writes for one reader and one subject. Each tell it looks for is a form of that default choice. It marks every tell it finds, strongest first, drafts a rewrite without treating the original structure as fixed, checks the draft against the patterns and the original claims, and then writes the final version. It explicitly does not invent facts: names, numbers, dates, quotes, and citations must come from the source or the writer, and when a needed detail is missing it asks rather than fabricating one. Output shows the work: first rewrite, a short critique of anything still artificial, and the final version. The 25 patterns are grouped into five sections and numbered by strength and frequency. The first five justify an edit on a single sighting; patterns marked "weak alone" count only when several tells share a passage. Section A covers staging instead of stating (not-X-but-Y constructions, one-line closers and dramatic fragments, sayings that sound deep, staged run-ups, arguing with no one). Section B covers rhythm by rule (forced triads, repeated sentence openings, dashes as universal connector, stacked qualifiers, hyphenated pairs everywhere, passive voice and missing subjects). Section C covers inflation and borrowed authority (overused AI words, inflated significance, vague connection or association, shallow -ing riders, sales language, borrowed authority, avoiding is/are/has). Section D covers formatting by rule (bold as decoration, decorative headings, curly quotation marks). Section E covers leftovers from the chat and draft (chatbot residue, knowledge-limit disclaimers and guesses, a heading repeated in the first sentence, writing about the previous version). The README includes a full before/after example: an AI-sounding Lisbon travel piece rewritten into a first-person account that keeps the original topic, view, and details while removing staged significance, sales language, and formulaic closers. The pattern list is sourced from Wikipedia's "Signs of AI writing" article, maintained by WikiProject AI Cleanup. The project is MIT licensed and has an extensive version history, with version 3.0.0 consolidating 35 patterns into 25 and realigning them with the current Wikipedia article.