Autonomous Dev
mingyang-arch/autonomous-dev
Desenvolvimento autônomo em loop: planejar → implementar → verificar → iterar
Ver no GitHubSkill para Claude Code que automatiza o desenvolvimento de software em ciclos fechados: planejamento, implementação, verificação e iteração. Funciona com qualquer stack (Node, Python, Rust, Go, etc.), detectando automaticamente os comandos de build e teste do projeto.
Instalação
npx convez-ai mingyang-arch/autonomous-dev1 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
- Implementação autônoma de features com testes até ficar verde
- Refatoração de código com verificação contínua
- Setup de projetos novos com gate de qualidade automatizado
Veredito

Excelente para quem quer delegar ciclos de desenvolvimento completos ao Claude. O princípio de "dar um check executável + um revisor independente" é sólido; funciona agnóstico de linguagem. Requer setup inicial de SPEC.md e CLAUDE.md, mas depois o loop roda sozinho.
README
A tradução para o português deste README ainda está na fila.
autonomous-dev
A language- and stack-agnostic Claude Code skill for driving the agent to build software autonomously over long stretches — plan, implement, verify, iterate — instead of one-shotting.
The core idea: a closed loop beats a longer prompt. The model stops when work looks done; give it a check it can run itself and an independent reviewer, and it iterates until the work is done.
Why stack-agnostic
It never assumes npm/pytest/cargo/go. It detects the project's own build and test commands first, then drives the same loop regardless of language — backend, frontend, CLI, data, infra, or libraries. Any codebase with a runnable check works.
What's inside
autonomous-dev/
├── SKILL.md # five-step workflow (detect → plan → gate → loop → review)
├── references/
│ ├── workflow.md # four phases + three autonomous-loop patterns
│ ├── verification.md # per-ecosystem command detection; what counts as a "check"
│ └── pitfalls.md # five failure modes and their fixes
├── assets/
│ ├── CLAUDE.md.template # persistent project rules
│ ├── SPEC.md.template # the contract: what + out-of-scope + how to verify
│ ├── todo.md.template # the ordered, checkable build loop
│ ├── check.sh # Stop-hook gate: auto-detects the test runner, blocks "done" while red
│ ├── settings.json # wires the Stop hook
│ └── reviewer-agent.md # adversarial fresh-context reviewer subagent
└── scripts/
└── init_autonomous_dev.sh # drops the gate + templates into any project
The Stop hook auto-detects the test command across npm / pytest / cargo / go / gradle / maven / mix / rspec / composer / make, and exits non-zero to block a turn from ending while tests fail.
Install
Clone into your Claude Code skills directory:
git clone https://github.com/mingyang-arch/autonomous-dev.git ~/.claude/skills/autonomous-dev
Use
Invoke it in Claude Code:
/autonomous-dev
…or just ask: "build X autonomously", "set up an autonomous loop", "implement this feature properly with tests".
To scaffold the gate + planning files into an existing project:
bash ~/.claude/skills/autonomous-dev/scripts/init_autonomous_dev.sh /path/to/project
Then fill in CLAUDE.md (the project's real commands), SPEC.md (goal + features), and todo.md (ordered steps), and start the loop.
The one principle
Give the model a check it can run, and let a different role verify the result.
Everything else is engineering detail around those two moves.
License
MIT — see LICENSE.