convez.ai
SkillEscrita & ConteúdoIntermediário
Heman10x-NGU

Unslop

Heman10x-NGU/unslop

Remove marcas de IA da prosa com proteção contra sobre-correção

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill que remove características típicas de escrita gerada por IA (frases genéricas, estruturas previsíveis, jargão corporativo) do texto mantendo o significado e autenticidade. Inclui guardrails para evitar over-correction que torna o texto artificial de outra forma.

Instalação

$npx convez-ai Heman10x-NGU/unslop

1 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

Estrelas
1
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Polir textos gerados por IA antes de publicação
  • Melhorar autenticidade de conteúdo em blogs, artigos e redes sociais
  • Treinar modelos a evitar padrões previsíveis de escrita
  • Avaliar qualidade de prosa em dimensões de naturalidade e densidade

Veredito

Skill bem estruturada que aborda o problema real de escrita AI genérica com fundamento teórico (raiz das falhas) em vez de apenas listas negras. A seção de guardrails é particularmente valiosa, evitando a armadilha comum de criar prosa robotizada ao tentar remover traços de IA.

README

A tradução para o português deste README ainda está na fila.

Unslop

A skill for removing AI tells from prose.

What this is

AI writing has patterns. Predictable phrases, structures, rhythms. This skill teaches Claude (or any LLM) to catch and remove them.

Design notes

  • Root-cause framing. The skill explains why models slop (they trade rare specific facts for common safe phrasing) so it generalizes past any fixed blacklist.
  • A guardrails section. Over-applied, de-slop rules grind prose into choppy fragments, no connective tissue, every sentence the same length, which is its own AI tell. The guardrails preserve meaning, allow justified passive voice, keep load-bearing adverbs, and allow a single em dash.
  • Wikipedia-catalogued tells. A reference file covering puffery, significance inflation, the "-ing" fake-analysis tail, copula avoidance, the "challenges and future" formula, sycophancy/leftover machine voice, and formatting tells, drawn from Wikipedia's ~15,000-word community catalog.
  • A hard gate. Never fabricate a fact to replace a vague one; never drop a real claim. Verify before scoring.

Skill Structure

unslop/
├── SKILL.md                  # Core instructions + guardrails
├── references/
│   ├── phrases.md            # Phrases to remove
│   ├── structures.md         # Structural patterns to avoid
│   ├── wikipedia-tells.md    # Puffery, significance inflation, copula avoidance, machine residue
│   └── examples.md           # Before/after transformations
├── README.md
└── LICENSE

Quick start

Claude Code: Add this folder as a skill.

Claude Projects: Upload SKILL.md and reference files to project knowledge.

Custom instructions: Copy core rules from SKILL.md.

API calls: Include SKILL.md in your system prompt. Reference files load on demand.

What it catches

Banned phrases - Throat-clearing openers, emphasis crutches, business jargon, all adverbs, vague declaratives, meta-commentary. See references/phrases.md.

Structural clichés - Binary contrasts, negative listings, dramatic fragmentation, rhetorical setups, false agency, narrator-from-a-distance voice, passive voice. See references/structures.md.

Sentence-level rules - No Wh- sentence starters, no em-dash overuse, no staccato fragmentation, no lazy extremes, active voice required.

Scoring

Rate 1-10 on each dimension:

DimensionQuestion
DirectnessStatements or announcements?
RhythmVaried or metronomic?
TrustRespects reader intelligence?
AuthenticitySounds human?
DensityAnything cuttable?

Below 35/50: revise.

License

MIT. Use freely, share widely.

Ver o original no GitHub

Bom para

Relacionados