qiaeru/skill-human-prose

writing-human-prose, a Claude Code skill that removes predictable AI writing patterns from English and French prose, or flags them without rewriting.
The skill treats two failure modes as slop. The first is the familiar AI gloss: filler phrases, formulaic contrasts, signature vocabulary ("delve," "leverage," "seamless"; « véritable », « permettre de », « implémenter »), vague declaratives, and pull-quote bait. The second is overcorrection: prose chopped into slogans, or made casual on purpose to "sound human," which reads as machine-made as the gloss. The target sits between them. The tells differ by language, so the rules do too: English prose is pushed toward fluent, natural US English with concrete subjects and varied rhythm, while French prose is pushed away from the short sentences calqued on English toward sentences articulated by subordination, with French typography.
The name says what the prose should sound like. The skill aims at no detector score: it counts the fake typo and the bolted-on casualness among the tells, and in audit mode it names and quotes patterns without guessing who wrote a text.
Whatever the language, the rewrite stays faithful to the source: it adds no number, cause, or actor the original doesn't contain, changes only what a check flags so the author recognizes the result as their own draft, and flags a vague passage to the author rather than fill it in.
## Credits
This skill was inspired by several other skills, including stop-slop by Hardik Pandya, humanizer by Siqi Chen, no-ai-slop by Peter Yang, anti-ai-slop-writing by jalaalrd, SlopMonster, anti-slop by Miqdad Badjuber, avoid-ai-writing by Conor Bronsdon, Not-Ai by udaysharmadev, anti-ai-writing-kit by evelynyaxueke, and anti-ai-writing-claude-skill by Avik Bal.
## Layout
text
skill-human-prose/
├── .claude-plugin/
│ ├── plugin.json
│ └── marketplace.json
├── .github/
│ ├── scripts/
│ │ ├── normalize.mjs
│ │ └── validate.mjs
│ ├── workflows/
│ │ └── validate.yml
│ ├── dependabot.yml
│ └── FUNDING.yml
├── evals/
│ └── - /
│ ├── prompt.md
│ └── graders/
├── .gitattributes
├── .gitignore
├── .markdownlint-cli2.jsonc
├── README.md
├── CHANGELOG.md
├── LICENSE
└── skills/
└── writing-human-prose/
├── SKILL.md
└── references/
├── en/
│ ├── rules.md
│ ├── phrases.md
│ ├── structures.md
│ ├── examples.md
│ └── ui-strings.md
└── fr/
├── rules.md
├── tournures.md
├── structures.md
├── typographie.md
├── interfaces.md
└── exemples.md
SKILL.md holds everything that doesn't depend on the language: the process, the output contract, voice and register, the weighing of tells, and the final read. It sends Claude to the rules file of the text's language, en/rules.md or fr/rules.md, which holds that language's default register, core rules, single-instance tells, and quick checks, and points to its catalogs and before-and-after examples. Each trigger loads the shared core and one language only. The French files are written in French, since their rules and examples are French.
The repository doubles as a Claude Code plugin named writing-human-prose and as its own plugin marketplace: plugin.json describes the plugin (the whole repository, with the skill under skills/), and marketplace.json lists it so Claude Code can install and update it straight from GitHub.