Gender Decoder para Anúncios de Emprego
POLFLAPS/claude-gender-decoder
Detecta e corrige linguagem codificada por gênero em job descriptions
Ver no GitHubSkill do Claude Code que detecta linguagem com viés de gênero em anúncios de vagas, combinando análise determinística baseada em pesquisa com revisão contextual via LLM. Identifica palavras codificadas (masculinas/femininas) e propõe reescritas neutras para aumentar inclusão.
Instalação
npx convez-ai POLFLAPS/claude-gender-decoder1 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
- Auditar anúncios de emprego para viés de gênero antes de publicar
- Melhorar inclusão em processos de recrutamento e atração de talentos
- Analisar múltiplas JDs em lote para conformidade de linguagem inclusiva
Veredito

Ferramenta sólida e baseada em pesquisa validada (Gaucher et al. 2011) que combina determinismo reproduzível com inteligência contextual do LLM. Útil para RH e recrutadores que buscam eliminar barreiras linguísticas sutis, mas não substitui auditoria completa de inclusão (ignora ageismo, ableismo, etc).
README
A tradução para o português deste README ainda está na fila.
claude-gender-decoder
A Claude Code skill that checks job ads for subtly gender-coded language — and helps you fix it.
It's a hybrid of a deterministic scorer and an LLM:
decode.pyreproduces the validated word-list method from the research, so the verdict is reproducible and citable — the same JD always scores the same.- The LLM layers context on top: it strikes false positives the word list can't (negation, domain terms), catches coded phrasing the list misses ("rockstar", "work hard / play hard"), and writes a balanced rewrite.
Based on Gaucher, Friesen & Kay (2011), Evidence That Gendered Wording in Job Advertisements Exists and Sustains Gender Inequality. Masculine-coded wording ("competitive", "dominant", "driven") makes roles less appealing to women and lowers their sense of belonging; feminine-coded wording has little effect on men.
Install (as a Claude Code skill)
Clone into your skills directory:
git clone https://github.com/POLFLAPS/claude-gender-decoder.git \
~/.claude/skills/gender-decoder
Then in Claude Code:
/gender-decoder paste-or-point-at-your-JD
Claude runs the decoder, reports the verdict, reviews the flagged words in context, and offers a neutral rewrite.
Use the scorer on its own (no Claude required)
decode.py is plain Python 3, standard library only:
python3 decode.py path/to/job-ad.txt # human-readable + JSON
python3 decode.py --json path/to/job-ad.txt # JSON only
echo "We need a confident, competitive leader" | python3 decode.py
Example output:
VERDICT: strongly masculine-coded
Masculine-coded hits : 5
Feminine-coded hits : 0
Coding score (F - M) : -5
Masculine-coded words found:
competitive, confident, ambitious, leader, dominate
Verdict scale
| Coding score (feminine − masculine) | Verdict |
|---|---|
| both 0 | no coded words |
| 0 (with hits) | neutral |
| +1 … +3 | feminine-coded |
| > +3 | strongly feminine-coded |
| −1 … −3 | masculine-coded |
| < −3 | strongly masculine-coded |
Masculine-coded is the verdict that measurably deters candidates — prioritise fixing those.
Caveats
- Binary framing — the research only studied masculine vs feminine coding and excluded non-binary participants/language. Directional signal, not a full inclusivity audit.
- "Neutral" ≠ bias-free — says nothing about ageism, ableist phrasing, or unnecessary requirements. Read the whole ad.
- English-only word list — it will mis-score non-English JDs.
- Stem matching is crude — which is why the LLM context-review step matters.
Tests
python3 tests.py
Credits & license
- Research: Gaucher, D., Friesen, J., & Kay, A. C. (2011), J. Personality & Social Psychology, 101(1), 109–128.
- Word lists & algorithm: adapted from the open-source Gender Decoder by lovedaybrooke (MIT, © 2016 lovedaybrooke), live at gender-decoder.katmatfield.com.
MIT licensed — see LICENSE. The license carries both copyright
notices.