anti-ai-slop
tazhate/anti-ai-slop
Claude Code skill that strips the LLM smell out of prose (EN + RU). Kills AI word-tells, rule-of-three, em-dashes, performative honesty.
Instalação
npx convez-ai tazhate/anti-ai-slop1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
anti-ai-slop
A writing skill for Claude Code that strips the LLM smell out of prose. Works for English and Russian.
You know the smell. "Let's delve into this multifaceted landscape." Every sentence the same length. Em-dashes everywhere. Three adjectives where one would do. The text is grammatical, polished, and unmistakably written by a machine.
This skill is a checklist against all of that. It loads automatically when Claude writes anything a human will actually read - an article, a README, an email, a PR description, a landing page. Short chat replies and code are left alone.
What it catches
The obvious stuff first: a kill-list of words that show up 10-50x more often in AI text than in human writing (delve, leverage, utilize, seamless, robust, and the Russian equivalents). Then the harder, structural tells that newer models still fall for:
- The rule of three. LLMs pack everything into triplets - "fast, reliable, and easy". The skill forces uneven list lengths.
- Empty participial tails. "...highlighting its importance", "...reflecting broader trends". Fake analysis glued to the end of a sentence. Cut it.
- Flat rhythm. Real writers mix long sentences with three-word ones. The skill tracks burstiness and sentence openers.
- Performative honesty. "Let's be honest", "no fluff", "here's the thing". Announcing directness instead of being direct. And the negative-parallelism tic: "it's not X, it's Y".
- Em-dashes and the Russian letter
ё- both stripped, per the author's house style.
The core idea, borrowed from Colin Gorrie's analysis: none of these devices are bad on their own. A tricolon is a classic rhetorical move. What gives the machine away is robotic consistency - the same trick in every paragraph. So the rule is almost never "never". It's "not back-to-back, not on autopilot".
Before / after
Slop: This robust framework leverages a comprehensive set of tools to streamline your workflow, highlighting its importance in today's fast-paced world.
Human: The framework ships with a test runner and a linter. That's it. No config, so you start writing in about a minute.
Install
Drop the skill into your Claude Code skills directory:
git clone https://github.com/tazhate/anti-ai-slop.git
mkdir -p ~/.claude/skills/anti-ai-slop
cp anti-ai-slop/SKILL.md ~/.claude/skills/anti-ai-slop/
Claude Code picks it up on the next session. It triggers on its own when you ask for any long-form text, or you can invoke it by name.
You can also lift the principles straight out of SKILL.md and paste them into a system prompt for any other model - nothing here is Claude-specific.
По-русски
Скилл двуязычный. Для русского у него свои таблицы калек и канцелярита (осуществлять -> делать, в связи с -> из-за), отдельный список показной честности ("если честно", "будем откровенны", "не просто X, а Y") и две привычки автора: короткий дефис вместо длинного тире и буква е вместо ё. Принципы те же, что в английской части - рваный ритм, конкретика, позиция вместо реферата, никаких триколонов на автомате.
Весь текст правил - в SKILL.md.
License
MIT. See LICENSE. Use it, fork it, paste it into your own prompts - do whatever.