convez.ai
SkillDevOps & InfraestruturaIntermediário
nhattrung0911

Shipwright — Disciplina de Engenharia para Agentes IA

nhattrung0911/shipwright

Ciclo completo: plan → build → test → prove → secure → operate

Ver no GitHub
InicialPluginScore · de 0 a 100

Framework de skills compostos que transforma agentes IA de codificação (Claude Code, Codex, Gemini) em engenheiros disciplinados. Implementa ciclo completo: planejar requisitos, cortar em slices testáveis, validar com testes, provar conclusão com evidência, aplicar gates de segurança OWASP Top 10:2025 e configurar controles de operação (rate limits, monitoring, backups). Elimina saltos diretos para código, testes faltando e afirmações falsas de conclusão.

Instalação

$npx convez-ai nhattrung0911/shipwright

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
0
Forks
0
Acompanhando
0
Saúde
57%
Último commit
há 4 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Estruturar desenvolvimento de dashboards SaaS com pilares de produção
  • Construir sistemas RAG com gates de segurança e testes integrados
  • Validar que código está realmente pronto antes de deploy com evidências
  • Implementar controles de dia-2 (monitoring, rate limiting, rollback)

Veredito

Skill robusto que ataca um problema real: agentes IA que fingem conclusão sem disciplina. A abordagem de ciclo de vida completo (planejar → construir → testar → provar → securizar → operar) é madura e baseada em repositórios canônicos de 50k+ ⭐. Excelente para times que querem elevar padrão de software gerado por IA, com suporte multi-modelo pronto.

README

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

Shipwright

⚓ Shipwright

Production-grade engineering discipline for AI coding agents

5 composable skills that make your AI agent plan, build, secure, and operate software like a senior engineer — never skipping a step, never faking "done".

License: MIT Skills Claude Code Also works on PRs welcome


❓ Why Shipwright — the problem it solves

AI coding agents are fast, but undisciplined. Out of the box they:

Without Shipwright ❌With Shipwright ✅
Jump straight to code, no planPlan first — requirements → testable slices
Write code, then maybe a testTest each slice before moving on
Say "Done! ✅" with no proof"Done" requires evidence — the test output that proves it
Forget security until the breachOWASP Top 10:2025 gate before shipping
Never think about staying alive in prodDay-2 controls — rate limits, monitoring, backups, rollback
Bloat the context, burn tokensToken-frugal — delegate, search-don't-read

The result without it: work that looks finished and isn't. Shipwright fixes the process, not just the prompt — it installs the instincts a senior engineer carries, as skills your agent loads automatically when they matter.

If your agent has ever confidently shipped something broken, this is the fix.

⚡ What you get

Plan first → cut into testable slices → test every slice → prove "done" with evidence → ship it secure → keep it running.

A coherent suite covering the whole lifecycle — Build → Secure → Ship → Operate. Stack-agnostic, distilled from canonical 50k+ ⭐ engineering repos and OWASP Top 10:2025, written in neutral prose so it runs on Claude Code, Codex, and Gemini CLI.

🚀 Install (Claude Code)

claude plugin marketplace add nhattrung0911/shipwright
claude plugin install shipwright@shipwright

That's it — skills auto-activate by context. No config.

🔄 How it works

flowchart LR
    A([You describe<br/>a task]) --> B{Shipwright<br/>activates}
    B --> C[1 · Frame<br/>requirements +<br/>acceptance criteria]
    C --> D[2 · Plan<br/>cut into<br/>testable slices]
    D --> E[3 · Build each slice<br/>test first → code → verify]
    E --> F[4 · Prove done<br/>evidence, not<br/>assertion]
    F --> G([Secured &<br/>operable])
    style B fill:#7c5cff,color:#fff
    style F fill:#34d399,color:#062

The engine (disciplined-delivery) runs this loop and pulls the other skills in as depth — security checks, ops gates, web pillars — exactly where each belongs.

📖 Usage guide (simple)

You don't call the skills manually — you just work. Describe what you want; the right skill fires on its own.

You say…Skill that firesWhat it does
"Build me a SaaS dashboard"shipping-production-websites + disciplined-deliveryFrames it, plans slices, walks 14 production pillars
"Build a RAG chatbot"disciplined-delivery + securing-applicationsPlan + test gates + prompt-injection / LLM checks
"Is this secure before I ship?"securing-applicationsOWASP Top 10:2025 gate with a runnable check per risk
"How do we keep it running?"operating-production-servicesRate limits, monitoring, backups, rollback, DR
"Refactor this large repo"token-frugal-engineeringKeeps context lean, delegates verbose work

Prefer to invoke one explicitly?

/shipwright:disciplined-delivery
/shipwright:securing-applications

That's the whole workflow: install once → describe tasks normally → the agent now plans, tests, secures, and proves its work automatically.

🧭 The five skills

SkillWhat it enforces
disciplined-deliveryThe engine. Frame → Plan → Slice → Verify for any project (web, LLM, ML, CLI, library, mobile, infra). Evidence-based "done", grader-≠-doer check.
shipping-production-websitesWeb: 12-layer full-stack map + 14 production pillars + pre-launch gate. Frontend, API, DB, auth, payments, i18n, SEO, config.
securing-applicationsOWASP Top 10:2025 → defense + a runnable check per risk. CSRF, SSRF, BOLA, JWT, file-upload, CORS, headers, LLM prompt-injection, data privacy.
operating-production-servicesDay-2: rate limiting, retries, circuit breakers, SLO, on-call, canary deploys, DR (RTO/RPO), backups + restore drills, cert rotation.
token-frugal-engineeringKeep the agent's context lean: delegate verbose work, search don't read, session hygiene.

🔗 How they chain

disciplined-delivery            ← the engine, every non-trivial build
├─ shipping-production-websites  ← pulled in for web projects
├─ securing-applications         ← the security gate (OWASP Top 10:2025)
├─ operating-production-services ← runtime controls + Day-2 maintenance
└─ token-frugal-engineering      ← cross-cutting, keeps context cheap

Use one, use all — they cross-reference but each stands alone.

✨ What makes it different

  • Evidence over assertion. "Done" requires the test output / eval score that proves it. Hedging words ("should work", "probably") are red flags the skill catches.
  • No step skipped. Pressured to "just ship it"? The plan shrinks to 3 lines — it never disappears.
  • Grader ≠ doer. A fresh check grades the work against criteria with a per-criterion PASS/FAIL + evidence table. Faking requires fabricating output.
  • Battle-tested sources. Distilled from system-design-primer, developer-roadmap, nodebestpractices, clean-code-javascript, awesome-scalability (all 50k+ ⭐) and OWASP Top 10:2025.

🌐 Works on Claude Code, Codex & Gemini

Skill bodies use neutral prose, so they port across agents.

AgentHow to install
Claude Codeclaude plugin marketplace add nhattrung0911/shipwright → claude plugin install shipwright@shipwright
CodexCopy the skills/* folders into ~/.agents/skills/
Gemini CLICopy the skills/* folders into ~/.gemini/skills/

📚 Browse the skills

🤝 Contributing

PRs welcome — new skills, sharper checks, fixes. See CONTRIBUTING.md. House rules: neutral/portable prose, evidence-based gates, lean bodies.

⭐ Like it?

If Shipwright makes your agent ship better, star the repo — it helps others find it.

📄 License

MIT — free for any use, personal or commercial.


Ver o original no GitHub

Bom para

Relacionados