Slop Check
allanta8/slop-check
Audita conteúdo para detectar AI slop com scoring em 8 dimensões
Ver no GitHubSkill que audita conteúdo em redes sociais (X, LinkedIn, artigos) para detectar padrões de AI slop (conteúdo genérico gerado por IA). Avalia 8 dimensões de qualidade, identifica violações automáticas e fornece recomendações de reescrita com veredicto final (PASS, NEEDS REWRITE ou REWRITE).
Instalação
npx convez-ai allanta8/slop-check1 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
Casos de uso
- Validar posts antes de publicar em redes sociais
- Avaliar qualidade e autenticidade de artigos e threads
- Identificar padrões de conteúdo genérico ou gerado por IA
- Obter feedback estruturado sobre voz e originalidade do texto
Veredito

Ferramenta especializada e honesta para quem cria conteúdo em redes sociais. O framework é maduro (baseado em Awwra), atualizado para abril de 2026, e fornece feedback granular e acionável — mas exige integração manual no workflow. Recomendado para criadores que querem evitar o clichê genérico.
README
A tradução para o português deste README ainda está na fila.
slop-check
A Claude Code skill that audits content for AI slop patterns and scores it across 8 dimensions. Built on the Awwra anti-slop framework (last updated April 2026).
What it does
Paste any content — a tweet, reply, thread, or article — and get back:
- Every Tier 1 violation with the exact offending quote
- Pass/fail on 3 quality gates (adds value, takes position, stops scroll)
- Scores across 8 dimensions (1–10 each)
- Top 3 issues with actionable rewrite directions
- A final verdict: PASS, NEEDS REWRITE, or REWRITE
Install
npx skills add ./slop-check -y -g
Or drop the slop-check folder directly into your ~/.agents/skills/ directory.
Usage
/slop-check
Then paste the content you want audited. The skill will ask if you don't provide it upfront.
You can also specify the platform:
/slop-check [X post]
/slop-check [ViewFT article]
/slop-check [LinkedIn post]
If not specified, the skill infers the platform from content length and style.
Scoring
| Dimension | What it measures |
|---|---|
original_take | Says something non-obvious |
no_ai_tells | Free of banned vocab, phrases, formatting |
voice_match | Sounds like a specific human, not a newsletter |
platform_fit | Appropriate for the target platform |
keyword_density | Low jargon-to-insight ratio |
no_engagement_bait | No hollow CTAs or "Thoughts?" closers |
no_emotional_inflation | No hype vocabulary or overclaiming |
humanness | Chaotic variation, rough patches, genuine texture |
no_ai_tells and humanness are weighted 1.5x in the overall score.
Thresholds:
- ≥ 8: PASS
- 6–7: NEEDS REWRITE (fixable)
- < 6: REWRITE (don't post as-is)
What it checks
Tier 1 — Automatic disqualifiers
Any single violation here means rewrite before posting.
- Banned formatting: em dash, bold markdown, emoji closers
- Banned openers: sycophantic agreement, restating the original post
- Banned filler phrases: "It's worth noting", "At the end of the day", "Furthermore,", etc.
- Banned vocabulary (3 layers):
- Original: pivotal, robust, delve, tapestry, synergy, certainly, etc.
- 2026 general: game-changer, move the needle, narrative, in this space, etc.
- April 2026 crypto/AI: execution risk, price action, conviction, flywheel, moat, macro tailwinds, etc.
- Banned structural patterns (3 layers):
- Original: symmetric takes, three bullet points, generic conclusions, engagement bait, etc.
- 2026 general: "we're early" safe harbor, authenticity performance, definition opener, uniform sentence weight
- April 2026: hourglass close, even paragraphing, template paragraph structure, definition stacking, policy voice without specifics, false certainty, consensus-adjacent take, asymmetric polish, fake burstiness
Tier 2 — Quality gates
- Gate 1: Adds at least one specific data point, named example, reframe, prediction, or first-person observation
- Gate 2: Takes a falsifiable position — someone smart could push back
- Gate 3: First 8 words earn the read — no context-setting openers
The core test
"Would a sharp human who actually lives in this space, with strong opinions, say exactly this?"
If the answer is "probably not" or "it's fine" — that's a 5. Fine is slop.
April 2026: second-generation slop
Newer models (GPT-5.1, Claude 4) suppress the obvious tells — em dash, "delve", three bullet points — by default. The slop hasn't gone away. It's gone underground.
The patterns to watch now:
- Fake burstiness — models trained to vary sentence length produce mechanical short-long-short-long alternation. Real human variation is chaotic. Read the draft aloud: if it sounds like a metronome, rewrite it.
- Asymmetric polish — intro and outro are fluent, middle is thin. Uniform polish throughout is itself a texture tell.
- Consensus-adjacent take — sounds opinionated but reflects the median view. AI defaults to the broadly acceptable middle ground.
- Even paragraphing — all paragraphs the same length. Humans don't distribute information symmetrically.
- Hourglass close — every post ends with a zoom-out synthesis. Humans often just stop when they've said the thing.
The suppressed-tell trap: a post conspicuously free of classic tells + second-gen patterns present is itself a signature. CT readers have calibrated for this since 2024.
Platform notes
- X / Twitter: Short posts under 100 chars skip even-paragraphing and template structure checks.
- ViewFT articles:
humannessandoriginal_takeweighted higher — the citation economy (ChatGPT, Perplexity, Grok) requires specificity, not just slop-free text. - Threads: First tweet scored for scroll-stopping; full thread scored for structural patterns.
- Replies: Gate 1 requires something the original tweet didn't already have.
Credits
Built by Viewfin Labs — the team behind ViewFT and Awwra.
- ViewFT — credibility-first publishing with ViewCred score
- Awwra — AI creator growth platform for crypto/AI/SaaS creators
Anti-slop framework developed for the @A11anTa and @viewftcom content pipelines.
License
MIT