Thoth — Skills de Agentes em Produção
taleshagn-horushub/thoth
Skills agentic produtivos para Claude e SDK Anthropic — do mundo real
Ver no GitHubBiblioteca aberta de skills agentic (system prompts, referências, scripts e templates) testadas em ambientes produtivos reais. Compatível com Claude, Anthropic SDK, GitHub Copilot e Cursor, transformando agentes genéricos em especialistas de domínio específico.
Instalação
npx convez-ai taleshagn-horushub/thoth1 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
- Especializar agentes IA para tarefas jurídicas e operacionais
- Criar estratégias de marketing automatizadas com contexto de negócio
- Orquestração de agents com prompts e guardrails consolidados
- Reutilizar conhecimento agentic entre Claude, Copilot e Cursor
Veredito

Abordagem pragmática ao problema real de escalabilidade de agents: prompts validados em produção, não em playgrounds. Ideal para equipes que rodam operações críticas com IA e precisam de reproducibilidade e segurança.
README
A tradução para o português deste README ainda está na fila.
Thoth
Skills your AI agents actually need.
Production-tested agentic skills, prompts, and processes — born in real businesses, not in playgrounds.
What is Thoth?
Thoth is a curated, open-source library of agentic skills — bundles of system prompts, references, scripts, and templates that turn a generic AI agent into a domain specialist.
Every skill in this repo:
- ✅ Is production-tested in a real business workflow before landing here
- ✅ Follows the Anthropic Skills format — drop-in compatible with Claude Code, the Agent SDK, and the Anthropic API
- ✅ Auto-converts to GitHub Copilot prompts/instructions and Cursor rules via the build pipeline
- ✅ Ships with usage examples, guardrails, and a clear definition of when not to invoke it
Named after Thoth — the Egyptian god of writing, knowledge, and magic. The skill library that pairs with the Horus ecosystem of AI-native legal operations.
Why this exists
Generic LLMs are smart but unfocused. The same model that writes a legal contract one minute and a sales email the next will produce mediocre results on both unless you scaffold it with the right context, references, and constraints.
That scaffolding is a skill — and writing good ones is hard, slow, and expensive. Most teams reinvent the wheel for every new use case.
Thoth is the inverse: a curated library where every skill has already proven itself in production, with the prompts, examples, and guardrails that make it work.
The mantra: if it hasn't been used in a real business, it doesn't belong here.
Quickstart
Use a skill in Claude Code
git clone https://github.com/taleshagn-horushub/thoth.git
cp -r thoth/skills/horus-content-creator ~/.claude/skills/
The skill is now available in any Claude Code session.
Use a skill via the Anthropic SDK
import Anthropic from "@anthropic-ai/sdk";
import { readFileSync } from "fs";
const skill = readFileSync("./thoth/skills/horus-content-creator/SKILL.md", "utf8");
const message = await new Anthropic().messages.create({
model: "claude-opus-4-5",
max_tokens: 4096,
system: skill,
messages: [{ role: "user", content: "Crie uma campanha Meta Ads para escritório trabalhista." }],
});
Use as a git submodule (recommended for production)
If your runtime needs to read skills at execution time (e.g. an event-driven orchestrator), pin them as a submodule:
git submodule add https://github.com/taleshagn-horushub/thoth.git vendor/thoth
This is how the horus-core orchestrator consumes Thoth skills in production.
What's inside
thoth/
├── skills/ # Anthropic Skills format (canonical source)
│ ├── horus-content-creator/ # Estratégia de marketing jurídico de alta conversão
│ └── market-research-offers-strategy/ # Wide research + Value Ladder
├── scripts/
│ └── build.ts # Generates Copilot/Cursor formats from canonical
├── docs/
│ ├── skill-format.md # Authoring spec
│ └── contributing.md # How skills get accepted
├── .github/workflows/
│ └── validate-skills.yml # CI: lint frontmatter, broken refs, etc
├── LICENSE # MIT
└── README.md
Skill format
Every skill is a folder with this layout:
my-skill/
├── SKILL.md # Required. System prompt + frontmatter (name, description)
├── references/ # Optional. Markdown files the skill reads on demand
├── templates/ # Optional. Output templates (markdown, txt, html)
└── scripts/ # Optional. Helper scripts (Python, Node, etc)
The frontmatter follows the Anthropic Skills spec:
---
name: my-skill
description: One paragraph that lets the model decide *when* to invoke this skill. Be specific.
---
See docs/skill-format.md for the full authoring guide.
Skills available
| Skill | Domain | What it does |
|---|---|---|
horus-content-creator | Marketing jurídico | Estratégia completa Meta Ads + Google Ads + Landing Page com compliance OAB |
market-research-offers-strategy | Estratégia comercial | Wide Research de competidores + criação de Value Ladder produtizada |
llm-council | Decision support | 5 AI advisors com lentes opostas + peer review anônimo + chairman synthesis. Adaptado da metodologia Karpathy |
perfil-executivo | Document generation | Gera Sumários / Perfis Executivos em PDF de alto impacto via WeasyPrint (HTML+CSS), com capa, sumário, encarte de oferta e branding |
🚧 More skills landing soon. See the roadmap.
Roadmap
The next batch of skills targets the 7 operational flows of an AI-first agency, including:
onboarding-strategist— diagnóstico inicial de cliente novobriefing-extractor— extração estruturada de briefings em texto livrecalendar-scheduler— agendamento via linguagem naturalmessage-composer— comunicação cliente em pt-BR mercado jurídicofollow-up-watcher— detecção de leads parados em pipelinecampaign-optimizer— análise diária de campanhas + ajustes propostosvideo-editor-brief— geração de roteiro técnico para edição
If your business has a workflow that an AI specialist could own end-to-end, open an issue — we want to hear it.
Contributing
Quality > quantity. Every skill in main has earned its place by surviving real production use.
If you want to contribute a skill, the bar is:
- Used in production for at least 30 days, by you or your team
- Self-contained — frontmatter, references, and templates all present
- Guardrailed — explicit rules for when not to invoke + how to refuse
- Tested — at least 3 example prompts with expected behavior in the PR description
See docs/contributing.md for full guidelines.
License
MIT — use freely, fork freely, just don't strip attribution from contributed skills.