Agent Skills
mattdweigand-sketch/agent-skills
Skills portáteis para agentes: eval loops, research, escrita e limpeza de projetos
Ver no GitHubBiblioteca portátil de skills reutilizáveis para agentes de IA (Codex, Claude Code, Cursor e similares). Cobre fluxos de avaliação, pesquisa, escrita, auditoria de código e higiene de projetos com 16 skills prontos para uso.
Instalação
npx convez-ai mattdweigand-sketch/agent-skills1 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 e otimizar alocação de lógica em projetos (60-30-10)
- Melhorar prompts e workflows através de teste iterativo (autoresearch)
- Refatorar grandes bases de código com múltiplas worktrees em paralelo
- Validar se sistemas de julgamento fecham loops entre padrões capturados e resultados reais
- Humanizar prosa gerada por IA para soar mais natural
- Definir e auditar estrutura limpa de repositórios
Veredito

Conjunto robusto de skills bem documentado para quem trabalha com agentes de IA em fluxos de desenvolvimento. Excelente para equipes que precisam padronizar e auditar workflows de agentes, embora exija familiaridade com arquitetura de prompts e lógica de avaliação.
README
A tradução para o português deste README ainda está na fila.
Agent Skills
Portable skill library for repo work, eval loops, research, writing, and agent workflow hygiene. These skills are written to work across Codex, Claude Code, Cursor, and other AGENTS-aware tools.
Included Skills
Each skill keeps its full workflow in skills/<skill-name>/SKILL.md.
| Skill | Description |
|---|---|
60-30-10 | Evaluate where a project keeps its judgment, and whether the allocation matches the durable shape: mostly owned data, partly deterministic code, only a thin layer of live model. |
autoresearch | Improves a prompt, config, or workflow by testing one change at a time against a fixed score. |
context-window-audit | Finds startup context, settings, wrappers, and memory entries that waste context-window space. |
cyborg-check | Audits whether a judgment system closes the loop between captured patterns and real outcomes, or is spreading unverified judgment. |
humanizer | Rewrites prose so it sounds clearer, more natural, and less AI-generated. |
karpathy-guidelines | Keeps coding work simple, scoped, and easy to verify. |
n-agentic-harnesses | Helps design or review the system around AI agents, tools, memory, permissions, and evals. |
parallel-refactor | Coordinates a large refactor across multiple isolated worktrees. |
ralf-loop | Runs a repeatable test-and-fix loop until the work passes, blocks, or needs a decision. |
repo-structure-audit | Defines and audits a clean repo layout for code and knowledge projects. |
research-synthesis | Turns a source set into clear findings, contradictions, gaps, and next questions. |
quality-check-eval | Sorts AI output quality checks into four buckets: automate in code, judge with an LLM, keep for human review, or remove. |
run-kit | Turns fuzzy agent work into a scoped run, then helps review or cross-check the result. |
runtime-guard | Prevents an agent from editing or publishing without clearly defined authority, boundaries, stop conditions, or rollback. |
skill-creator-v3 | Creates, improves, tests, and packages reusable skills. |
skill-tune | Audit or refactor a skill or prompt artifact for prompt technical debt. |
weekly-signal-diff-ai | Summarizes what structurally changed in AI this week and why it matters. |
Install
Copy all skills into the skill directory used by your agent:
cp -R skills/* ~/.agents/skills/
For Codex, use:
cp -R skills/* ~/.codex/skills/
Install one skill by copying only that folder:
cp -R skills/ralf-loop ~/.agents/skills/ralf-loop