AI text has tells: inflated wording, filler, the rule of three, empty enthusiasm, chatbot artefacts. This plugin removes them and leaves plain, precise prose, in three languages:
| Skill | Language | Language-specific layer |
|---|---|---|
humanizer-en |
English | English word lists and typography, legal-register mode, post-edit fact check |
humanizer-pl |
Polish | Polish slop vocabulary, calques from English, Polish typography, 2026 spelling reform, legal-register mode, post-edit fact check |
humanizer-pt |
Portuguese (pt-BR default, pt-PT on request) | gerundism, pronominal "mesmo", calques, variant mixing, legal-register mode, post-edit fact check |
All three share one taxonomy, built on Wikipedia's "Signs of AI writing". Claude picks the edition that matches the text. Each skill instructs Claude to rewrite wording and keep the facts; review the result as you would any edit.
Each edition ships one opinionated default: a single kind of dash (hyphen only) and a calm, precise voice. If you give Claude your own style guide, a writing sample or a voice skill, yours wins. The anti-slop patterns apply either way.
A quality pass for prose. Whether and how you disclose AI use stays your call; the skills make the writing better and leave that decision to you.
In Claude Code:
/plugin marketplace add matematicsolutions/matematic-humanizer
/plugin install matematic-humanizer@matematic-humanizer
Single-skill zips, no plugin needed, are on the MateMatic Boutique: English · Polish · Portuguese
The skills have no connectors and make no network calls of their own. They work on the text in your conversation, which Claude processes the same way as any other message you send it.
Each edition ships one local check (Python standard library only): post_edit_audit.py
(English), audyt_po_przepisaniu.py (Polish) and auditoria_pos_edicao.py (Portuguese). It
compares the text before and after editing and reports whether dates, amounts, legal
references, the force of obligations and legal terms survived. It reads the two files you
point it at and nothing else, writes nothing and opens no connections. Claude Code asks before
running it; where scripts cannot run, the skill does the same check by hand.
MIT. Adapted from blader/humanizer (MIT); see
NOTICE for full attribution.