ai-slop-cleaner-en-ru

An AI-slop cleaner and detector for Claude Code / the Claude Agent SDK.
A bilingual (EN + RU) skill: above all it cleans the AI out of text, rewriting
clichés into natural prose while preserving meaning, facts and terms — and it also
diagnoses "AI-likeness" (how machine-written a passage is, by which markers, with
what confidence).

Demo

Screen text:

"When it comes to performance, caching is not just useful, it's essential. Let's
dive in. Moreover, it scales. In conclusion, a seamless experience." 🚀

→ Almost certainly AI (~95/100). Tells: "not just… it's", "let's dive in",
the "moreover / in conclusion" cluster, a decorative emoji, zero specifics.

Clean text (register-aware):

before: Moonshot just dropped a new model, you can spin it up locally.
after: Moonshot released a new model; you can run it locally.

Formal register → slang normalized, every fact left intact.

What it does

Give it an article, forum post, PR description, essay or comment and it:

rewrites the slop out into natural human prose, preserving every fact, number,
term, quote and code block — it never invents content;
or, on request, just screens the text: rates its "AI-likeness", lists the exact
markers it found, and honestly reports confidence. AI detection is unreliable —
translation, editing and plain good writers all cause false positives, and the
skill says so instead of pretending otherwise.

It works in both languages. English tells follow Wikipedia's "Signs of AI writing";
Russian tells cover everyday clichés plus the Habr/vc.ru tech-blog style ("под
капотом", "давайте разберёмся").

It detects the register first and rewrites within it. Official press post,
news, blog or forum comment — the skill figures out the tone up front and keeps it
consistent: it won't leave casual slang in a formal announcement ("выкатила" →
"представила"), and it won't flatten a lively forum comment into stiff officialese.
Register consistency is treated as its own axis, separate from the AI-vs-human one.

It also tells decorative AI emoji apart from genuine human emoji:

AI emoji = cold section markers prefixing headings/bullets (🚀 Launch, 💡 Key idea,
✅ Next steps), one "corporate" emoji per item — that's a tell, and it gets stripped
when cleaning;
human emoji = hot, reactive, in-sentence ("this is peak 🔥", "I'm dead 💀😂") — that
counts toward a human author.

Emoji on their own are never treated as slop. Scoring is a calibrated judgment, not
rigid point-math — the skill is deliberately not boxed into a fixed template.

What it catches

Russian clichés: «важно отметить», «не просто X, а Y», «в современном мире»,
«играет ключевую роль», vague «эксперты считают»;
Russian tech-blog slop (Habr/vc.ru): «под капотом», «давайте разберёмся»,
«разложим по полочкам», «и вот тут на сцену выходит»;
English tells (Wikipedia AI Cleanup): "not just… it's a testament", "delve",
"pivotal role", AI-vocabulary clusters;
structure: the rule of three, uniform rhythm, boilerplate
"Intro / Conclusion / Future prospects" sections, decorative emoji markers,
empty upbeat endings;
and human counter-signals (typos, slang, personal numbers, argument, replies
to other people, live reactive emoji) that lower the score.

What the output looks like

There is no fixed form — the skill answers in plain language and adapts to the text
and the question (a yes/no gets a one-liner; "break it down" gets a full pass).
Roughly:

Almost certainly AI (~95/100), medium confidence.
Decisive: "the cache isn't just storage, it's a whole acceleration mechanism"
(negative parallelism) + a double "under the hood" with zero specifics. Plus
"hard to overstate the role of" and a 🚀 emoji heading. Nothing human: no numbers,
no first-hand experience. No counter-signals.

The number is a ballpark, not a precise figure, and confidence is stated honestly
(short text → low confidence).

Install

Claude Code (as a user skill)

git clone https://github.com/tuwulalo/ai-slop-cleaner-en-ru.git ~/.claude/skills/ai-slop-cleaner-en-ru

On Windows:

git clone https://github.com/tuwulalo/ai-slop-cleaner-en-ru.git "$env:USERPROFILE\.claude\skills\ai-slop-cleaner-en-ru"

Restart Claude Code — the skill is picked up automatically and triggers when you ask
to clean AI out of a text or to check a text for AI.

As a project skill (for a team/repo)

Drop the folder into .claude/skills/ai-slop-cleaner-en-ru inside your project and
commit it — the skill becomes available to everyone working in the repository.

Usage

Just ask in plain language:

"Was this written by a neural network?" + paste a passage
"Clean the AI out of this" / "rewrite this like a human, kill the slop"
"Is this AI?" + a link to an article
"Screen every .md in docs/ for slop and give me a table"
«Вычисти ИИ из этого текста» / «проверь, не нейросеть ли это писала»

Modes: diagnose one text (full report), clean (rewrite + change log + new score),
batch over files (table + breakdown of the worst), quick yes/no, or by URL (via
WebFetch).

Layout

ai-slop-cleaner-en-ru/
├── SKILL.md                  # skill instructions (detect + clean), the emoji rule
├── reference/
│   ├── markers-ru.md         # Russian markers (general + tech-blog)
│   ├── markers-en.md         # English markers (Wikipedia AI Cleanup)
│   ├── human-signals.md      # human counter-signals
│   ├── scoring.md            # scoring anchors and verdict bands
│   └── rewrite.md            # cleaning mode: "cliché → natural replacement" tables
├── examples/
│   └── examples.md           # annotated samples for calibration
├── scripts/
│   ├── markers.mjs           # shared regex/heuristic for the tools (not the skill)
│   └── scan.mjs              # batch triage: marker density across files
├── eval/
│   ├── cases.json            # labeled snippets (ai / human / mixed)
│   └── run.mjs               # regression guard over the marker set
├── validate.mjs             # spec validation for SKILL.md (no Python needed)
└── CHANGELOG.md

Development & validation

Repo tooling — no dependencies, Node 18+ (the skill itself needs none of this):

node validate.mjs            # check SKILL.md against the Agent Skills spec
node eval/run.mjs            # regression-guard the marker lists on labeled cases
node scripts/scan.mjs docs/  # triage your own files by marker density

validate.mjs mirrors what skills-ref
checks (name/description limits, field formats, referenced files) but needs no Python.
eval/run.mjs runs a deterministic marker heuristic (scripts/markers.mjs),
not the model. It guards the marker lists: if an edit breaks AI-vs-human
separation on the labeled set, it fails. The subtle "mixed" class (translation,
edited AI) is intentionally left to the model.
scripts/scan.mjs is triage, not a verdict — point it at your content.
Pointed at this repo's own reference/ it will flag those files, because the
marker dictionaries literally quote the clichés they describe.

What it does NOT do

It doesn't "prove" authorship — it gives a probabilistic verdict with a confidence level.
When cleaning it invents no facts and doesn't touch numbers, terms, quotes or code —
it removes clichés, not content.
It doesn't replace judgment: on short text (<120 words) confidence is always low.

Sources of the patterns

Wikipedia: Signs of AI writing (WikiProject AI Cleanup)
Википедия: Признаки сгенерированности текста
Breakdowns of ChatGPT/DeepSeek clichés on vc.ru, otzyvmarketing, Skillbox; observations of the Habr/vc.ru style.

License

MIT — see LICENSE.