Verificador de Fatos para Redes Sociais
hodoshia/social-media-fact-checker
Detecta alucinações de IA em posts antes de publicar
Ver no GitHubSkill que valida automaticamente cada afirmação factual em rascunhos de posts de redes sociais antes da publicação, detectando alucinações de IA (estatísticas falsas, leis, fontes, citações inventadas). Separa fatos de opiniões, verifica contra fontes primárias e retorna versão corrigida com fontes citadas.
Instalação
npx convez-ai hodoshia/social-media-fact-checker1 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
- Validar posts com dados/estatísticas antes de publicação
- Verificar conformidade regulatória e precisão legal em comunicações
- Agências e equipes de comunicação reduzirem risco de desinformação
Veredito

Ferramenta robusta e necessária para equipes que publicam em redes sociais e não podem arriscar erros factuais. Abordagem rigorosa contra alucinações (verifica fontes primárias, hierarquia de confiabilidade) é rara e bem-executada; caso de uso crítico para compliance e credibilidade.
README
A tradução para o português deste README ainda está na fila.
social-fact-check
A Claude Code skill that fact-checks a social media post draft before you publish it.
Paste a draft (LinkedIn, Facebook, Instagram, X/Twitter, TikTok, ...) and ask for a fact check. The skill extracts every claim, verifies each against authoritative sources, flags anything wrong, unverified, hallucinated, or overstated, and hands back a corrected, source-backed version in the brand's own voice.
Built for teams (agencies, comms, social) who cannot afford to publish a wrong fact, and for the reality that a lot of social copy is now drafted by AI.
Why: AI hallucinations
Language models write a lot of social copy now, and they hallucinate. They invent statistics, laws, dates, document numbers, sources, and quotes, then state them with total confidence. The output reads perfectly and is sometimes simply false.
Confidence is not evidence. This skill treats every factual claim in a draft as potentially fabricated until a real, primary source confirms it. It checks that cited sources actually exist and actually say what the draft claims, instead of trusting a plausible-looking URL or a real-sounding number.
Posting a wrong figure, a misquoted law, or a "proposal" stated as "approved" is a real credibility and legal risk, especially for regulators and public bodies. This skill makes the check systematic instead of a gut feeling.
What it does
- Infers the client/brand and its domain from the current project context (no hardcoded client).
- Separates FACT claims from EDITORIAL/opinion.
- Verifies each fact adversarially via web search, using a strict source-tier hierarchy (primary law and official records first; blogs and interested-party releases never count alone).
- Runs anti-hallucination checks: confirms cited sources, studies, laws, and document numbers actually exist and actually say what the draft claims; an unverifiable "confident" statistic is treated as fabricated.
- Applies accuracy rules: proposal vs decision, stage vs in-force, who proposes vs who decides, exact terminology, no overstated causation, ranges for contested numbers.
- Returns a verdict table, the required changes, a cleared version of the post, and the residual risks.
Install
This repo is a Claude Code plugin and its own marketplace. Inside Claude Code:
/plugin marketplace add hodoshia/social-media-fact-checker
/plugin install social-fact-check@social-media-fact-checker
/reload-plugins
Verify with /plugin list. The skill is then available as social-fact-check.
Manual install (no plugin system)
Copy the skill folder into your Claude Code skills directory instead:
git clone https://github.com/hodoshia/social-media-fact-checker /tmp/sfc \
&& cp -r /tmp/sfc/skills/social-fact-check ~/.claude/skills/ \
&& rm -rf /tmp/sfc
Restart Claude Code (skills load at session start).
Use
Paste your draft and ask, for example:
Fact-check this LinkedIn post before I publish it: ...
or invoke it directly:
/social-fact-check
The skill picks up the brand and voice from your project context (e.g. a CLAUDE.md, brand or voice guide). The corrected version comes back in the draft's language and the brand's tone.
Example
Draft pasted in (AI-written LinkedIn post):
AI adoption is exploding. A recent Gartner report found that 87% of companies will fully replace their support teams with AI by 2026. That is why our platform is the only way to stay competitive. Book a demo today.
What the skill returns:
Verdict table
| # | Claim (quote) | Type | Verdict | Source |
|---|---|---|---|---|
| 1 | "a recent Gartner report found that 87% of companies will fully replace their support teams with AI by 2026" | FACT | WRONG / hallucinated | No such Gartner figure exists. Tier 1-2 search finds no 87% or "fully replace by 2026" claim. Classic fabricated citation. |
| 2 | "our platform is the only way to stay competitive" | EDITORIAL | OPINION (overstatement) | "the only way" is unsupported. |
| 3 | "AI adoption is exploding" | EDITORIAL | OPINION (acceptable) | Vague but fine as framing. |
Required changes
- Remove the fabricated Gartner statistic, or replace it with a real, cited figure.
- Soften "the only way" to a defensible claim.
Cleared version
AI adoption is accelerating, and support teams are a common starting point. Our platform helps you keep up without rebuilding your stack. Book a demo today.
Residual risks
- If you want a stat in the post, supply a real source and the skill will verify the exact number and wording before it goes back in.
This is the core value: the 87% Gartner line reads perfectly and is simply invented. The skill catches it before it reaches your audience.
Customize
- Point it at your brand voice: keep a voice/style guide in the project so the cleared version matches your tone.
- Adjust strictness: it weights rules up automatically for regulators and public bodies; you can tell it the brand type.
- Higher confidence: it can run multiple independent verification passes on high-stakes claims.
License
MIT. See LICENSE.