convez.ai
SkillEscrita & ConteúdoIntermediário
ojesusmp

Voz

ojesusmp/voz

Escreva como humano, mate sinais de IA, ajuste tom ao leitor

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill para Claude Code que remove marcas características de escrita gerada por IA e adapta o tom do texto a diferentes públicos. Usa análise baseada em Wikipedia e estudos de corpus para gerar prosa clara, honesta e humanizada, com nove registros profissionais intercambiáveis.

Instalação

$npx convez-ai ojesusmp/voz

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
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 1 mês
Versão
—
Issues abertas
1
Contribuidores
1

Casos de uso

  • Remover linguagem padronizada de IA (delve, tapestry, robust, seamless)
  • Adaptar tom para vendas, legal, marketing, documentação técnica ou jornalismo
  • Gerar prosa clara e confiante em registros específicos (empático, corporativo, técnico, motivacional)

Veredito

Ferramenta focada e bem-documentada que aborda um problema real: textos que soam robotizados. Vai além de detectar IA, orienta para escrita humanizada. Requer integração manual com Claude Code, viável para quem já usa o ecossistema.

README

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

voz

A Claude Code skill that makes writing read like a person wrote it, and lets you switch voice to fit the reader.

This is writing guidance, not an AI detector. The goal is clear, honest prose. It is never a tool for labeling, scoring, or accusing anyone's text, and it should not be used that way.

voz ("voice" in Spanish) does two things:

  1. Kills AI-writing tells. The overused words (delve, tapestry, underscore, boasts, robust, seamless), the rule-of-three, the "it's not X, it's Y" pattern, the inflated "serves as a" instead of "is", and the formatting overkill (bold-label bullet lists, em dashes, emoji headers, Title Case Headings). Strip those and text stops sounding machine-made.
  2. Switches register on demand. One warm, plain, confident default voice, plus eight professional registers you can call by name.

It is designed to run by default once you wire it into your own Claude Code setup (a pointer in your CLAUDE.md, or a SessionStart hook). The repo itself is just the skill; the always-on wiring is yours to add. You should not have to ask for good writing.

The registers

RegisterFor
Empathetic (default)docs, support, chat, most things
Salesman / closer1:1 selling, proposals, objection handling
Corporateboard updates, all-hands, investor notes
Lawyer / legalcontracts, demand letters, terms, notices
Journalistannouncements, news, press-release rewrites
Marketinglanding pages, ads, headlines, email
Technicalspecs, APIs, runbooks, dev docs
Teacherguides, onboarding, explainers
Motivationalpep talks, vision, closers (use sparingly)

Say the register's name and the writing shifts. Say nothing and you get the default.

What's in here

SKILL.md                  the spine: default voice, kill-list, register table, self-check
references/
  ai-tells.md             full tell catalog, with sources and dates
  registers.md            all nine register profiles (diction, moves, sample openings)
  human-prose.md          the positive craft: specificity, rhythm, restraint
  self-check.md           a 20-second pass to run before you deliver
scripts/
  voz_sessionstart.py     the lightweight reminder the always-on hook prints
  wire_settings.py        idempotent always-on wiring (CLAUDE.md + settings.json)
install.ps1               one-command installer (Windows)
install.sh                one-command installer (macOS/Linux)

Install

Quick install

Clone the repo and run the installer. It copies the skill into ~/.claude/skills/voz, which both the Claude Code CLI and the VSCode extension read.

Windows (PowerShell):

git clone https://github.com/ojesusmp/voz.git
./voz/install.ps1

macOS/Linux (bash):

git clone https://github.com/ojesusmp/voz.git
bash voz/install.sh

Always-on (optional)

A plain install just adds the skill, so you invoke it when you want it. To make the default voice apply automatically in every session, add the always-on flag. It appends a pointer to your global ~/.claude/CLAUDE.md and a small SessionStart hook to ~/.claude/settings.json. Both edits are idempotent and touch nothing else in your config.

./voz/install.ps1 -AlwaysOn
bash voz/install.sh --always-on

Always-on needs Python on your PATH, since the hook is a short Python script. If Python is missing the installer skips the wiring and tells you, leaving the skill itself installed.

Manual install

If you would rather not run a script, copy the folder yourself:

git clone https://github.com/ojesusmp/voz.git ~/.claude/skills/voz

Verify

Reload VSCode (Developer: Reload Window) or start a new Claude Code session, then ask Claude to "write a short product description and avoid AI tells", or name a register: "rewrite this in the journalist register."

How it was built

The avoid-list is grounded in real sources, not vibes:

The nine register profiles were each researched against style references (Bryan Garner for legal, AP and Reuters for journalism, HBR for corporate, classical rhetoric for motivational, and more, cited inline in references/registers.md).

Two honest caveats, baked into the skill:

  • The word list ages. "delve" spiked in 2023 and faded by 2025. The list is datable and meant to be refreshed. The patterns outlast the words.
  • No single tell is proof. One em dash means nothing; the "em dash equals robot" claim is overstated. The skill judges clusters, and it is writing guidance, not a detector. The aim is to write well, not to accuse anyone's text.

License

MIT. See LICENSE.

Ver o original no GitHub

Bom para

Relacionados