convez.ai
SkillAutomaçãoAvançado
newmindsgroup

Biblioteca Unificada de Skills para Agentes IA

newmindsgroup/ai-agent-skills-library

Skills reutilizáveis versionadas para agentes IA em múltiplas plataformas

Ver no GitHub
InicialPluginScore · de 0 a 100

Repositório centralizado de skills versionadas e revisadas por código para agentes IA, compatível com Claude Code, Cursor, Codex, Windsurf e outras plataformas via instalador universal. Implementa especificação aberta Agent Skills com suporte a runbooks, padrões operacionais e marketplace cross-IDE.

Instalação

$npx convez-ai newmindsgroup/ai-agent-skills-library

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
2
Forks
0
Acompanhando
0
Saúde
57%
Último commit
há 1 mês
Versão
—
Issues abertas
1
Contribuidores
1

Casos de uso

  • Construir biblioteca corporativa de skills padronizadas para múltiplas IDEs
  • Centralizar e versionar rotinas de automação para agentes IA
  • Compartilhar padrões operacionais e runbooks entre equipes

Veredito

Infraestrutura robusta e bem-pensada para governança de skills em organizações. Excelente para times que precisam escalar agentes IA com patterns consistentes, mas requer investimento inicial em adoção e manutenção do repositório.

README

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

AI Agent Skills Library

A versioned, code-reviewed library of agentic AI skills for the team. Works across Claude Code, Cursor, Codex, Windsurf, OpenCode, and Google Antigravity.

Every skill follows the open Agent Skills specification — a single SKILL.md plus optional references/, scripts/, and assets/ — and stays portable across organizations through a project-local brand-config.yml.

Why this exists

AI models are commodity. The harness (skills, memory, runbooks, coordination) is the moat. This repo is our Dojo: a code-reviewed, versioned marketplace of skills + patterns + installers so one person's breakthrough becomes everyone's baseline.

What's inside

.
├── .claude-plugin/marketplace.json   ← Claude Code plugin marketplace manifest
├── agents/                            ← 16 sub-agents (cherry-picked from msitarzewski/agency-agents, MIT)
│   ├── ai-engineer.md, sre.md, code-reviewer.md, incident-response-commander.md, ...
│   └── README.md                       ← install + auto-routing guide
├── sources/                           ← canonical source tree grouped by source repo
│   ├── original/skills/               ← New Minds-created skills
│   ├── sickn33-antigravity-awesome-skills/skills/
│   └── menkesu-awesome-pm-skills/skills/
├── dist/skills/                       ← generated flat Agent Skills export for installers
├── runbooks/                          ← non-loadable patterns + lessons
│   ├── gate-pattern.md
│   ├── systemd-watchdog-patterns.md
│   ├── drive-clone-handoff.md
│   ├── defensive-api-bindings.md
│   ├── read-before-escalating.md
│   ├── agent-operating-principles.md
│   └── tool-leverage-heuristics.md   ← when to fire which tool (the autonomy playbook)
├── installers/                        ← bootable kits
│   └── superpowers-stack/             ← Superpowers + 5 official MCPs + agency-agents (16) + Firecrawl + Graphify
├── prompts/                           ← reusable system-prompt templates
│   ├── voice-dna-template.md
│   ├── agent-skills-spec-summary.md
│   └── institutional-ai-stance.md
├── docs/
│   ├── INSTALL.md                     ← per-IDE install detail
│   ├── AUTHORING.md                   ← how to write a new skill
│   ├── PLACEHOLDERS.md                ← brand-config.yml placeholder system
│   ├── SKILLS-CATALOG.md              ← full index w/ dependencies + decision tree
│   ├── STARTER-PACKS.md               ← curated subsets for easier installs
│   └── SKILLS-TRUST.md                ← provenance, trust levels, and risk flags
├── manifests/
│   ├── skills-manifest.json           ← generated canonical manifest
│   └── starter-packs.json             ← curated bundle definitions
├── CHANGELOG.md                       ← versioned history of every release
├── brand-config.example.yml           ← copy to project root: tenant config, voice/visual paths
├── DESIGN.example.md                  ← copy to project root: visual SSOT (Stitch 9-section format)
├── AGENTS.example.md                  ← copy to project root: brand-context for any AI tool (open standard)
├── install.sh / install.ps1           ← cross-IDE one-liner installer
└── scripts/                           ← validation + scaffolding helpers

Quick install

Option A — Claude Code plugin marketplace (recommended for Claude Code users)

# In Claude Code:
/plugin marketplace add newmindsgroup/ai-agent-skills-library
/plugin install agent-foundation@ai-agent-skills-library

The marketplace exposes five bundles:

  • agent-foundation — operating principles + design methodology + human-voice + brand-blueprint-builder + tool-leverage-heuristics + brand-ssot-precedence. Load on every project. Six skills make any AI tool act as an institutional agent: signal over noise, on-brand voice, strategic tool-fire decisions, brand canon precedence.
  • content-engine — newsletter, social, brand-drift.
  • ops-pipeline — inbox triage, calendar, pipeline, competitive intel.
  • agent-self-improvement — weekly meta-review that improves your CLAUDE.md.
  • agency-agents — 16 specialist sub-agents (engineering, orchestration, content, visual). Auto-route on natural-language triggers. Tenant-agnostic when installed via this bundle (defers to brand-ssot-precedence for voice).

For the full catalog with dependencies + install priority + decision tree, see docs/SKILLS-CATALOG.md.

Option B — Cross-IDE one-liner (Cursor, Codex, Windsurf, OpenCode, Antigravity)

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/newmindsgroup/ai-agent-skills-library/main/install.sh | bash

Windows PowerShell:

iwr -useb https://raw.githubusercontent.com/newmindsgroup/ai-agent-skills-library/main/install.ps1 | iex

The installer auto-detects your IDE and installs every skill into the right location. It supports:

  • Claude Code → .claude/skills/ or ~/.claude/skills/
  • Cursor → .cursor/skills/ (with .cursor/rules/ fallback)
  • Codex / AGENTS.md → generates an AGENTS.md at repo root pointing to skills/
  • Windsurf → .windsurf/skills/
  • OpenCode → .opencode/skills/ or ~/.config/opencode/skills/
  • Google Antigravity → .agent/skills/ or ~/.gemini/antigravity/skills/

Option C — Install a single skill

curl -fsSL https://raw.githubusercontent.com/newmindsgroup/ai-agent-skills-library/main/install.sh | bash -s -- newsletter-drafter

Option D — Install a starter pack

curl -fsSL https://raw.githubusercontent.com/newmindsgroup/ai-agent-skills-library/main/install.sh | bash -s -- --bundle codex-essentials

Useful starter packs include:

  • agent-foundation — small default baseline for every project.
  • codex-essentials — routing, planning, debugging, verification, current docs, and skill discovery.
  • builder-stack — architecture, API design, implementation, review, and validation.
  • frontend-ui — frontend, visual design reference, animation, accessibility, and UX.
  • business-strategy — business model, startup validation, strategy, OKRs, metrics, and AI product.
  • security-review — security audit, API security, SAST, cloud security, and agent-action auditing.
  • qa-debugging — dogfooding, browser QA, bug hunting, TDD, and completion verification.
  • local-ai-ops — local models, graphing, design references, Claude routing, and AI-native CLI work.

List packs any time:

curl -fsSL https://raw.githubusercontent.com/newmindsgroup/ai-agent-skills-library/main/install.sh | bash -s -- --bundles

Option E — Search before installing

curl -fsSL https://raw.githubusercontent.com/newmindsgroup/ai-agent-skills-library/main/install.sh | bash -s -- --search "business model"
curl -fsSL https://raw.githubusercontent.com/newmindsgroup/ai-agent-skills-library/main/install.sh | bash -s -- --category "Frontend, UI, UX, Design" --dry-run

Installer discovery and safety options:

  • --list lists all skills.
  • --search <query> searches names, descriptions, categories, sources, trust levels, and bundle memberships.
  • --bundles lists starter packs.
  • --categories lists generated categories.
  • --bundle <name> installs a curated starter pack.
  • --category <name> installs every skill in a category.
  • --dry-run previews the target and skill count without copying files.
  • --force or --update refreshes selected existing skills.
  • --all explicitly installs the full library.

Option F — Manual

Each installable skill export lives at dist/skills/<skill-name>/. Copy the folder into your tool's skill directory. See docs/INSTALL.md for the full per-tool path table.

Configure brand context

Some skills (newsletter, social, triage, drift-scan, pipeline) need brand-specific context — voice rules, content pillars, entity-separation rules, schedules. Drop a brand-config.yml at the root of your project. Start from brand-config.example.yml and fill in your values.

Without brand-config.yml, those skills will surface a "missing config" message rather than guessing or using personal data from a previous user.

Full placeholder reference: docs/PLACEHOLDERS.md.

Currently shipped

Skills

This repository now ships 1513 skills. Canonical source folders live under sources/<source-repo>/skills/, and the installer-ready flat export lives under dist/skills/.

For the complete generated index with descriptions, use docs/SKILLS-CATALOG.md. For provenance and duplicate-resolution details, use docs/SKILL-SOURCES.md. For curated starter packs, use docs/STARTER-PACKS.md. For trust metadata and risk flags, use docs/SKILLS-TRUST.md.

Source organization

The canonical source tree is sources/. It groups skills into repo-named folders such as sources/sickn33-antigravity-awesome-skills/skills/, sources/menkesu-awesome-pm-skills/skills/, and sources/original/skills/ for New Minds-created skills.

The flat dist/skills/ directory is generated for compatibility with tools that expect skills-root/<skill-name>/SKILL.md. Do not edit dist/skills/ directly; edit sources/ and run ruby scripts/sync-skill-library.rb.

Generated discovery artifacts live beside the export:

CategoryCount
AI, Agents, LLMs, Data735
Automation, Integrations, Productivity188
Backend, APIs, Databases97
Business, Strategy, Product, Growth121
Content, SEO, Writing, Docs56
DevOps, Cloud, Infrastructure59
Frontend, UI, UX, Design111
General, Workflow, Research, Miscellaneous62
Languages, Frameworks, Specialist Tools25
Media, Documents, Office, Creative5
Security, Compliance, Risk19
Testing, QA, Debugging, Review35

Curated marketplace bundles remain intentionally small and high-signal:

  • agent-foundation: institutional-ai-operating-principles, design-methodology, human-voice, brand-blueprint-builder, tool-leverage-heuristics, brand-ssot-precedence
  • content-engine: newsletter-drafter, social-content-drafter, voice-drift-scanner
  • ops-pipeline: inbound-triage, calendar-scheduler, pipeline-reporter, competitive-monitor
  • agent-self-improvement: self-improvement-review

Runbooks

See runbooks/README.md for the full table. Six patterns covering deploy gates, watchdog discipline, drive-handoff, defensive bindings, escalation hygiene, and operating principles.

Installers

See installers/README.md. Currently one: superpowers-stack, which installs the full agent baseline on a Linux VPS via a single client.env:

  • Superpowers plugin (engineering rituals: planning, TDD, systematic debugging, verification, code review)
  • 5 official MCP servers (Memory, Fetch, Filesystem, Playwright, Chroma) — Tier-1 always-on
  • agency-agents cherry-pick (16 sub-agents from msitarzewski/agency-agents MIT) — auto-route via "Use PROACTIVELY when..." description triggers
  • Firecrawl MCP — 14 tools for web scraping, search, structured extract, persistent browser sessions
  • Graphify — turns any folder of code/docs/PDFs/images into a queryable knowledge graph (/graphify slash command)

Run with bash installers/superpowers-stack/scripts/install-all.sh (idempotent — safe to re-run). Configure once via installers/superpowers-stack/config/client.env.

Brand-context drop-ins

Three top-level templates you copy to the root of any project where you want consistent agent behavior:

Together: voice rules + visual rules + agent context, version-controlled, agent-readable, deliverable to clients.

Prompts

See prompts/README.md. Three reusable system-prompt templates for non-skills environments (ChatGPT custom GPT, Gemini Gem, Claude Project, Cursor rule).

Contributing a new skill

  1. Read docs/AUTHORING.md — the full authoring guide.
  2. Copy sources/original/skills/institutional-ai-operating-principles/ as a starting template, or use the scaffold script: ./scripts/new-skill.sh drafting-proposals.
  3. Write your SKILL.md following the Agent Skills spec (under 500 lines, gerund naming, third-person description, examples at the bottom).
  4. If your skill needs brand context, document the required brand-config.yml keys at the top of SKILL.md and use {{PLACEHOLDERS}} in the body. See docs/PLACEHOLDERS.md.
  5. Regenerate the flat export and indexes: ruby scripts/sync-skill-library.rb.
  6. Validate locally: ./scripts/validate-all.sh.
  7. Check generated drift and links: ./scripts/check-generated.sh && ruby scripts/check-links.rb.
  8. Smoke-test installer behavior: ./scripts/smoke-install.sh.
  9. If validation warns about long SKILL.md files, move deep material into references/ or run ruby scripts/compact-long-skills.rb /tmp/validation.log 501 10000 against a saved validation log, then run the sync script again.
  10. Open a pull request. CI will re-validate every skill on push.

What this repo does not do

There is no true "universal auto-installer" protocol today. No spec exists in 2026 that lets every IDE take a Git URL and automatically figure out which adapter to use. This repo's install.sh is the closest practical equivalent: it reads the environment, detects the IDE, and installs the right files. The Claude Code marketplace path (Option A above) is the supported native path for Claude Code specifically.

Every skill is portable — the canonical SKILL.md is the Agent Skills standard, and adapters are thin pointers to it, not parallel copies.

License

MIT — see LICENSE.

Ver o original no GitHub

Bom para

Relacionados