convez.ai
SkillEscrita & ConteúdoAvançado
Julz7727

Stop Slop v3

Julz7727/Stop-Slop-v3

Roteamento positivo para prosa humanizada sem padrões genéricos de IA

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill para gerar prosa limpa e humanizada a partir de LLMs, usando roteamento positivo em vez de restrições negativas. Integra padrões estruturais de registros específicos e permite calibração por amostra de voz do usuário, evitando a geração de conteúdo genérico e artificial.

Instalação

$npx convez-ai Julz7727/Stop-Slop-v3

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
5
Forks
3
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
v3.0.0
Issues abertas
1
Contribuidores
1

Casos de uso

  • Produção de conteúdo editorial com voz consistente
  • Escrita técnica e ensaios com registro específico
  • Refinamento de saída de LLM para publicação profissional

Veredito

Abordagem sofisticada que resolve um problema real: em vez de listar o que evitar (ativando exatamente esses padrões), constrói uma bacia de patterns positivos onde o slop não existe. Exige compreensão técnica mas entrega controle fino sobre a voz do output.

README

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

Stop Slop v3

A skill for producing clean, human-register prose from LLMs. Builds on Stop Slop v2 (Wake & Claude, Axiom), which itself rebuilt Hardik Pandya's Stop Slop. v3 integrates the humanizer skill's Wikipedia-derived pattern catalogue into the editorial layer without compromising v2's architectural principle.

The Problem with v1

Stop Slop v1 is a restriction-prompting architecture: lists of phrases to avoid, structures to reject, and checks that scan for bad patterns. When a human editor reads these lists and then writes, the method works. When injected into an LLM's context window, the method backfires.

Loading "Here's the thing:" and "Let that sink in" and "Not because X. Because Y." into the model's context primes those exact tokens. The model orients around the patterns it is supposed to suppress and spends inference energy resisting them — often producing output that avoids the literal prohibited form but stays semantically adjacent to the same basin.

Negative prompting activates the basin it intends to suppress.

What v2 Did

v2 introduced positive routing constraints that make AI prose patterns structurally impossible without ever naming them. The mechanism has three parts:

1. Derivation before generation. Before writing, the model completes five fields: readerProfile, registerTarget, sceneAnchor, textureConstraint, and optionally antiPattern. This commits the model to a specific voice basin before the first token.

2. Six simultaneous routing constraints. Every sentence must have a human subject, a material anchor, single-pass assertion, rhythm variation, register fidelity, and earned emphasis. These are presence tests, not absence tests.

3. Positive register patterns. Instead of listing structures to avoid, v2 provides structural habits that belong to specific registers — technical, conversational essay, literary, business, fiction, instructional. The model adopts the habits of the target register, occupying a basin where slop patterns do not exist.

The diagnostic taxonomy from v1 was preserved as a human-facing editorial reference, explicitly separated from the generation context.

What v3 Adds

v3 keeps everything in v2 intact and adds three things on top:

1. Voice Sample (sixth derivation field, optional). When you provide a sample of your own writing, v3 extracts its structural features (sentence rhythms, word choices, openers, transitions) and adds them as a voiceSample field in the derivation. The basin to route into becomes your actual voice, not "the closest publication that resembles your voice."

2. Voice Within the Register. A new SKILL.md section that names the positive moves available inside a chosen register basin — holding a position, naming mixed feelings, using first person where allowed, permitting small structural mess, specifying feelings instead of labeling them. These were implicit in v2; v3 makes them options the model can reach for.

3. Final Audit Pass. An optional two-step post-generation editorial pass: "what makes this obviously AI?" followed by a targeted revision. Explicitly framed as the only place in the skill where the diagnostic taxonomy may enter context, and only on a finished draft.

v3 also expands diagnostic-taxonomy.md with a second part: the twenty-nine specific surface patterns from Wikipedia's "Signs of AI writing" article, each with before/after revisions. Part I (v2's eight Pandya-derived mechanism categories) is preserved unchanged.

Skill Structure

stop-slop-v3/
├── SKILL.md                            # Core routing instructions + v3 additions
├── README.md                           # This file
├── CHANGELOG.md                        # v2 → v3 diff
├── LICENSE                             # MIT, with v2 + humanizer provenance
└── references/
    ├── derivation.md                   # Six derivation fields (v3 adds voiceSample)
    ├── routing-patterns.md             # Positive patterns across six registers
    ├── examples.md                     # Four worked examples: restriction vs. routing
    └── diagnostic-taxonomy.md          # Part I: 8 mechanism categories.
                                        # Part II: 29 surface patterns (v3).
                                        # NOT for model injection.

Quick Start

Claude Code: Clone this folder as a skill.

mkdir -p ~/.claude/skills
git clone https://github.com/<your-username>/stop-slop-v3.git ~/.claude/skills/stop-slop-v3

Claude Projects: Upload SKILL.md, references/derivation.md, and references/routing-patterns.md to project knowledge. Do NOT upload references/diagnostic-taxonomy.md — it is for human review and the Final Audit Pass only.

Custom instructions: Copy the core routing constraints and derivation fields from SKILL.md and references/derivation.md.

API calls: Include SKILL.md and references/derivation.md in your system prompt. Load routing-patterns.md when the register family is identified. Load diagnostic-taxonomy.md only when running the Final Audit Pass on a finished draft.

Using the Voice Sample (v3)

To match your personal writing voice, provide a sample of your own writing alongside the task:

Use stop-slop-v3. Here is a sample of my writing for voice matching:
[paste 2-3 paragraphs]

Now draft a 500-word post about [topic].

The skill will analyze your sentence rhythms, word choices, punctuation habits, and recurring phrases, then populate the voiceSample field of the derivation so the generated prose routes into your actual voice rather than a generic register.

Architecture

Built on Retrieval-Shaped Prompting & Constrained Synthesis v3:

  • Code/systems handle combinatorics and hard state (not applicable for pure prose, but present in tools like NPC Forge and Wheel of Existence where this methodology originated)
  • Context/prompting handles routing and governance (derivation fields, register targeting, routing constraints)
  • Model acts as a bounded semantic resolver within the routed basin (generates prose under constraint)
  • Validation handles persistence and correction (scoring rubric, Final Audit Pass, human editorial review via the diagnostic taxonomy)

Attribution

  • Mechanism categories (Part I of the diagnostic taxonomy) derived from Hardik Pandya's Stop Slop (MIT). Reorganized and annotated with causal explanations in v2 by Wake, Claude (Axiom), Lisa, Eos, and Gemini across the Local Group research program.
  • v2 architecture and SKILL.md by Chaosman-One: Stop-Slop-v2.
  • Surface pattern catalog (Part II of the diagnostic taxonomy) drawn from Wikipedia: Signs of AI writing (CC BY-SA 4.0), via the humanizer skill by Siqi Chen (MIT).

License

MIT. See LICENSE.

Ver o original no GitHub

Bom para

Relacionados