convez.ai
MCP ServerDesenvolvimento WebIntermediário
arvindand

Agent Skills

arvindand/agent-skills

Skills portáteis para assistentes de código com inteligência de dependências e design UI

Ver no GitHub
BomPluginScore · de 0 a 100

Conjunto de skills reutilizáveis para assistentes de IA como Claude Code, GitHub Copilot e Cursor. Oferece inteligência de dependências, operações GitHub, busca de documentação e padrões de design UI/UX com ativação automática baseada no contexto.

Instalação

$npx convez-ai arvindand/agent-skills

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
17
Forks
2
Acompanhando
1
Saúde
42%
Último commit
há 2 semanas
Versão
v1.7.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Consultoria automática de documentação de bibliotecas (React, Next.js, Prisma)
  • Operações GitHub com análise profunda de repositórios
  • Análise de dependências JVM e planejamento de upgrades com detecção de CVEs
  • Criação de interfaces UI/UX acessíveis e funcionais
  • Geração de skills reutilizáveis a partir do histórico de sessões

Veredito

Ferramenta madura e bem arquitetada para potencializar assistentes de IA em fluxos de desenvolvimento. Excelente cobertura de casos de uso reais (docs, GitHub, dependências, UI), mas requer ecossistema compatible e configuração inicial.

README

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

Agent Skills

Portable skills for AI coding assistants. Primarily tested with Claude Code and leverages Claude Code-specific features (hooks, context forking), but core functionality works with any Agent Skills-compatible tool including GitHub Copilot, OpenCode, and Cursor.

Installation

Option 1: Install as Plugin (Recommended - For Claude Code)

# Add the marketplace
/plugin marketplace add arvindand/agent-skills

# Install the plugin
/plugin install agent-skills@arvindand-skills

Or test locally:

claude --plugin-dir /path/to/agent-skills

Option 2: Copy to Skills Directory

# Clone and copy skills
git clone https://github.com/arvindand/agent-skills.git
cp -r agent-skills/skills/* ~/.claude/skills/

Works with Claude Code, GitHub Copilot, OpenCode, Cursor, and VS Code with Copilot.

Available Skills

SkillDescriptionUse For
context7Library documentation lookup via Context7 REST APIGetting up-to-date docs for React, Next.js, Prisma, etc.
github-navigatorGitHub operations via gh CLI with deep analysis modeAll GitHub operations + codebase analysis via cloning
maven-toolsJVM dependency intelligence via Maven Tools MCP serverVersion checks, POM-aware upgrade planning, JVM dependency bot replacement flows, CVE scanning, license compliance
skill-craftingCreate, fix, validate skills + generate from session historyCreating skills, fixing issues, CSO compliance, session-to-skill conversion
ui-ux-designCreate production-grade interfaces with strong UX foundationsBuilding functional, accessible, visually distinctive UI/UX

Output Styles

Customize Claude's response style via /output-style:

StyleDescription
socraticTeaches through questions. Backs off when you ask for direct answers.
speed-runMax 3 sentences. Code first, explanations on request.
pair-programmerThinks aloud, invites interruption. Best for complex problems.
deep-focusCalm, methodical. Accuracy over speed. Best for hard problems.

Usage

Once installed, skills activate automatically when relevant to your prompt:

You: "How do I use React hooks?"
→ context7 fetches up-to-date React hooks documentation

You: "Show me open issues in facebook/react"
→ github-navigator uses gh CLI to list issues

You: "Analyze the architecture of vercel/next.js"
→ github-navigator clones repo for deep codebase analysis

You: "Should I upgrade Spring Boot from 2.7 to 3.2?"
→ maven-tools analyzes versions, CVEs, same-major safe fallback paths, and documentation handoff

You: "Build me a login form with dark mode"
→ ui-ux-design creates accessible component with proper states

You: "Create a skill from this session"
→ skill-crafting evaluates patterns and generates reusable skill

No manual invocation needed — the AI determines when each skill is relevant.

Documentation

See skill-crafting/REFERENCES.md for best practices and patterns.

Cross-Platform Design

Skills use progressive enhancement:

  • Core fields (name, description) work everywhere
  • Claude Code features should be ignored by other platforms

Claude Code Enhancements

When running in Claude Code, these skills leverage additional features:

FeatureSkillsWhat It Does
Context forkinggithub-navigatorRuns in isolated subagent to avoid polluting main context
Stop hooksskill-crafting, ui-ux-designVerifies task completion before declaring done
Tool restrictionsAllLimits which tools each skill can use

Other platforms get core functionality without these enhancements.

Contributing

Skills I'm looking to collect:

  • Frequent operations with zero context overhead
  • CLI tools that can be discovered via --help
  • Discovery patterns that teach AI dynamically

Note: Would appreciate contributions or references to implementations for other useful skills, especially geared toward helping senior devs focused on backend, architecture, and DevOps.

Why Skills over MCP?

I'm biased towards skills over MCP. Here's why.

Skills are cheaper and at least as effective as MCP tools when done well

MCP loads all tool schemas into every conversation whether you use them or not. Ten tools? That's roughly 1,000 tokens added to every single request. Update (2026): Anthropic's tool search tool now defers schemas and loads them on demand — but it's default-on only in Claude Code, opt-in elsewhere, and most MCP clients still preload. The overhead is reduced, not gone.

Skills are free until you need them. When a skill triggers, you pay for ~100 words of metadata. That's it.

Anthropic found that using code execution patterns (what skills enable) cut token usage from >150,000 to 2,000. That's a 98.7% reduction. https://www.anthropic.com/engineering/code-execution-with-mcp

If there's a CLI, use a skill

AI models already know how to read --help output. You don't need to write MCP schemas for things like gh, npm, or curl.

Instead, teach the pattern:

  • "Run gh issue --help to see what's available"
  • "Check npm --help for commands"

The skill stays current as the CLI evolves. No maintenance needed.

See Simon Willison, "Claude Skills are awesome, maybe a bigger deal than MCP" (Oct 2025) — the source of the CLI-over-MCP argument.

When the line is blurry, optimize for cost

Sometimes both approaches work. When in doubt, ask: will I use this frequently? If yes, a skill costs you nothing when idle. An MCP costs you tokens on every request.

When MCP makes sense

Use MCP when:

  • Works and well maintained and doesn't contain a gazillion tools
  • You need bidirectional communication (push updates, subscriptions)
  • Carries complex state and sophisticated caching
  • No CLI exists and you can't easily wrap the API

Skills and MCP can work together. You can write a skill that teaches the AI how to use your MCP servers effectively.

The maven-tools skill is a good example of that hybrid model: the MCP server does the dependency intelligence work, and the skill keeps the orchestration patterns lean, portable, and easier to evolve.

What's changed (2026)

The skills-vs-MCP split has aged toward convergence:

  • OpenAI adopted the Agent Skills standard in Codex and uses skills internally — so a single SKILL.md now travels across Claude Code, Copilot, Cursor, OpenCode, and Codex.
  • Anthropic donated MCP to the vendor-neutral Agentic AI Foundation (Linux Foundation, co-founded with OpenAI and Block).

Skills and MCP are becoming shared, cross-vendor infrastructure — which is why I reach for skills first but keep both.

License

MIT License — See individual skill LICENSE files for details.

Author: Arvind Menon


Ver o original no GitHub

Bom para

Relacionados