convez.ai
SkillEscrita & ConteúdoIntermediário
forint573

Sonnopus Humanizer

forint573/miAI-Humanizer-Skill-Awesome

Remove marcas de IA em copy de marketing, mantém voz autêntica

Ver no GitHub
BomPluginScore · de 0 a 100

Skill que humaniza textos de marketing e conteúdo longo gerado por IA, removendo marcas de processo e artificialismo. Remove referências ao chat, planos e rascunhos, preserva a voz da marca, nunca inventa fatos e mantém o foco no leitor e no produto.

Instalação

$npx convez-ai forint573/miAI-Humanizer-Skill-Awesome

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
13
Forks
2
Acompanhando
0
Saúde
85%
Último commit
há 4 meses
Versão
v1.0.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Polir landing pages e páginas de vendas geradas por IA
  • Refinar e-books, case studies e relatórios longos
  • Remover referências ao processo de criação e rascunhos
  • Preservar voz de marca em conteúdo humanizado

Veredito

Ferramenta prática e bem-focada para profissionais de conteúdo que usam IA e precisam remover sinais de processo. Abordagem honesta: não tenta enganar detectores de IA, apenas melhora legibilidade e foco do texto.

README

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

The Sonnopus Humanizer ✸☽

Humanize AI-written marketing and long-form copy in Claude. It removes AI tells and process bleed, keeps your facts and your voice, and never narrates the process behind the work.

CI License: Apache 2.0 Claude Skill Made for Sonnet & Opus


TL;DR

The Sonnopus Humanizer drafts and de-slops marketing and long-form copy: landing pages, e-books, reports, case studies, sales pages, product copy, and founder notes. It keeps the writing about the reader and the product, it refuses to invent facts, and it preserves a real voice instead of flattening it into house style.

It is built for copywriters, marketers, and founders who draft with AI and want copy that does not read like AI. It is not an AI-detector bypass. It makes weak copy read like a person actually wrote it.

Install it in one line, then ask Claude to humanize your copy.

git clone https://github.com/forint573/miAI-Humanizer-Skill-Awesome.git && \
  cp -r miAI-Humanizer-Skill-Awesome/the-sonnopus-humanizer ~/.claude/skills/

Then open Claude and say "humanize this landing page" or "finalize this draft and strip the planning notes." That is the whole setup.


What it is

The Sonnopus Humanizer is an Agent Skill for marketing and long-form copy. It fixes one common failure: the finished piece starts narrating the hidden work behind it. The copy points at the chat, the outline, the roadmap, an earlier draft, or a leftover scaffold instead of speaking to the reader about the product, the offer, the proof, the objection, and the next step.

It also avoids the two ways humanizing usually backfires:

  • It never invents numbers, names, testimonials, results, awards, or claims. Missing proof becomes a visible placeholder, not a fabrication.
  • It never flattens good human writing into one generic voice. A provided writing sample beats the skill's default taste every time.

The one test

Every sentence in the output should pass this question:

Would this sentence make sense to a reader who never saw the project, chat, plan, outline, prompt, or drafting process?

If not, it is process bleed. Cut it, or rewrite it around the reader, product, offer, proof, objection, or next step.

See it in action

The left cell carries two problems at once: it narrates the chat and the plan, and it ships an unverified number as fact. The right cell is what the skill returns.

BeforeAfter
"Based on our discussion, this guide starts with the problem, then moves into the framework we landed on. Our platform cuts onboarding time by 73% and is trusted by leading enterprises.""Onboarding breaks down in three predictable places. Here is where, and how to fix each one. [ADD VERIFIED METRIC] [ADD NAMED CUSTOMER]"

The chat reference and the roadmap narration are gone. The copy opens on the reader's problem. The unsupported 73% and the vague "leading enterprises" are flagged for a human to verify, not shipped as proof.

Install

Three ways, fastest first.

1. One line, Claude Code. Installs a personal skill that is available in every project:

git clone https://github.com/forint573/miAI-Humanizer-Skill-Awesome.git && \
  cp -r miAI-Humanizer-Skill-Awesome/the-sonnopus-humanizer ~/.claude/skills/

For a project skill that travels with one repo, copy the folder into that repo's .claude/skills/ instead.

2. Installer script. From a clone, run ./install.sh for user scope, or SCOPE=project ./install.sh to install into the current repo. To install without cloning first:

curl -fsSL https://raw.githubusercontent.com/forint573/miAI-Humanizer-Skill-Awesome/main/install.sh | bash

3. Packaged .skill file. Download it from the Releases page, or build it yourself:

make build   # writes dist/the-sonnopus-humanizer.skill

Upload that file anywhere custom Claude skills are accepted. The archive keeps the skill folder at its root.

Use it

Once installed, Claude reaches for the skill on its own when you ask for things like:

  • "Humanize this landing page and make it sound less like AI."
  • "Finalize the homepage copy from the outline. Remove planning notes and make it paste-ready."
  • "Draft a sales page for this offer using only the proof in the brief."
  • "Clean up this e-book chapter. Keep the arguments, but lose the self-announcing intros."

It stays out of the way for code, raw data, tables, changelogs, release notes, legal terms, technical reference docs, and casual chat. There, factual structure matters more than this posture.

Optional scanner

For long drafts, a heuristic scanner flags leftover process bleed, wrapper text, hype clusters, vague authority, and em or en dashes:

python the-sonnopus-humanizer/scripts/copy_scan.py draft.txt
# or
cat draft.txt | python the-sonnopus-humanizer/scripts/copy_scan.py

It produces false positives on purpose. Treat each result as a reason to look, not a verdict. Human judgment always wins.

This README was written with the skill's own rules: no em dashes, no hype clusters, no process bleed, no invented facts. The one place the rules are broken on purpose is the Before example above, which exists to show the patterns the skill removes.

What's inside

the-sonnopus-humanizer/
├── SKILL.md                      # the active instruction layer
├── README.md                     # the skill's own package notes
├── references/
│   ├── process-bleed.md          # the seven leaks, examples, fast-scan phrases
│   ├── ai-tells.md               # diagnostic catalog of generic AI prose
│   ├── qa-scorecard.md           # 8-category, 0 to 2 readiness score
│   └── voice-calibration.md      # matching a sample or brand voice
├── scripts/
│   └── copy_scan.py              # optional heuristic scanner
└── tests/
    └── test-prompts.md           # trigger / no-trigger checks and samples

Design notes

  • Lean active layer. SKILL.md holds only what the model needs in context to act. The deeper catalogs live in references/ and load on demand, which keeps the instruction layer clear for both Sonnet and Opus.
  • Diagnose on clusters, not words. ai-tells.md names why something reads as machine written. One formal word, one transition, or one dash is not proof of AI writing.
  • Lightest effective edit. Drafting, finalization, and cleanup are separate modes. The skill picks the smallest one that solves the task instead of rewriting good prose for its own sake.

FAQ

Who is this for? Copywriters, marketers, founders, and anyone who drafts marketing or long-form copy with AI and wants it to read like a person wrote it.

Does it bypass AI detectors? No. This is not an AI-detection bypass tool. It improves real copy: it removes AI tells and process bleed, flags unverified claims, and keeps your voice. The goal is honest, publishable writing, not gaming a detector.

Which models does it work with? It is written for Claude and tuned for both Sonnet and Opus. It works anywhere Claude reads skills, including Claude Code, the desktop and web apps, and any setup that accepts custom skills.

Will it invent facts to fill a gap? No. Missing proof becomes a visible placeholder such as [ADD VERIFIED METRIC], never a fabricated number, name, or testimonial.

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md and the Code of Conduct. Changes to behavior should keep tests/test-prompts.md honest and pass the automated checks (make check).

License

Apache License 2.0. Copyright 2026 Virág Làzár Csaba ✸☽. See LICENSE and NOTICE.

"Claude", "Sonnet", and "Opus" are model names from Anthropic. This is an independent community skill. It is not affiliated with or endorsed by Anthropic.

Ver o original no GitHub

Bom para

Relacionados