convez.ai
SkillAutomaçãoAvançado
eruto-skills

Autopilot

eruto-skills/autopilot

Sessões autônomas com critérios de aceição, orçamento e proteção contra alucinações

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill do Claude Code que scaffolda sessões autônomas de longa duração com critérios de aceitação observáveis, memória de lições entre iterações e parada automática em ambiguidades. Ideal para tarefas multi-etapas que você deixa rodando sem supervisão contínua.

Instalação

$npx convez-ai eruto-skills/autopilot

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
50%
Último commit
há 1 mês
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Automação de fluxos multi-etapa sem supervisão constante
  • Tarefas de longa duração com limites de custo e iterações definidos
  • Processamento com memória persistente de aprendizados entre ciclos

Veredito

Skill maduro e bem documentado para quem precisa orquestrar loops autônomos com segurança. Exige compreensão clara de critérios de aceitação observáveis e condições de parada — não é bala de prata para qualquer tarefa repetitiva.

README

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

autopilot

A Claude Code skill that scaffolds long-running, hands-off sessions without burning money on the wrong problem.

What it does

/autopilot <task> walks you through:

  1. Plan — define the task as a list of subtasks, each with a user-observable acceptance criterion
  2. Budget — set hard caps on iterations, cost, and wall-clock
  3. Mode — pick interactive (/loop), local cron, or cloud (/schedule Routines)
  4. Hand-off — print the exact command to start the loop

The loop itself is built on top of Anthropic's official /ralph-loop plugin and the bundled /loop skill. autopilot adds three regimens those alone don't enforce:

  • The agent re-reads the plan + accumulated lessons at the top of every iteration (deterministic stack — context never bloats)
  • Three consecutive failures on a task halts the run (truth-bias defense)
  • Ambiguity surfaces as <status>NEEDS_HUMAN</status> instead of a guess

Install

Clone the repo somewhere stable, then run the installer for your OS. It links the skill into ~/.claude/skills/autopilot/ so Claude Code discovers it. Idempotent — safe to re-run.

git clone https://github.com/eruto-skills/autopilot.git
cd autopilot
# Windows (no admin / dev-mode needed — uses a directory Junction)
pwsh -File scripts/install.ps1
# macOS / Linux
bash scripts/install.sh

Then in any Claude Code session: /autopilot <task description>.

If you can't run the script

The installer is just one command. Manual equivalent:

# Windows
New-Item -ItemType Junction `
  -Path  "$env:USERPROFILE\.claude\skills\autopilot" `
  -Target (Resolve-Path .)
# macOS / Linux
ln -s "$(pwd)" ~/.claude/skills/autopilot

File layout

autopilot/
├── SKILL.md                       # the skill definition (workflow Claude follows)
├── README.md                      # this file
├── templates/
│   ├── PLAN.md                    # task list with user-observable acceptance tests
│   ├── RUN.md                     # loop body — fed to Claude every iteration
│   ├── LESSONS.md                 # cross-iteration memory (Reflexion-style)
│   └── BUDGET.md                  # caps + halt conditions
└── references/
    ├── completion-criteria.md     # the single highest-leverage page — read first
    ├── halt-conditions.md         # exit statuses + the truth-bias problem
    ├── modes.md                   # interactive / local-cron / cloud comparison
    └── prior-art.md               # which pattern came from where

When to use vs not use

Use when: multi-step task, you'll walk away, mistakes are expensive to undo, you want a postmortem.

Don't use when: single repetitive poll (use /loop directly), already-clear one-shot prompt (use /ralph-loop directly), or you'll be at the keyboard anyway.

The non-obvious pieces

If you only read three things in this repo:

  1. references/completion-criteria.md — why "tests pass" is not an acceptance criterion
  2. references/halt-conditions.md — why 3 failures is the rule, not 5 or 10
  3. templates/RUN.md — the actual loop body; everything else exists to make this small

Prior art

This skill composes ~20 patterns from Anthropic's official docs, the English Ralph-loop community, the Japanese 自走 community, and agentic-loop research. Full attribution in references/prior-art.md.

License

Inherits the parent repo's license.

Codex / Claude Code installation

For an existing clone, run bash scripts/install.sh --host codex on macOS/Linux, or ./scripts/install.ps1 -HostName codex in Windows PowerShell. The installer keeps Claude as its default target and preserves existing directories.

This package supports both Codex and Claude Code. The plugin entry point is skills/autopilot/SKILL.md; the root SKILL.md remains the standalone source.

For Codex, add the public eruto-skills marketplace in the plugin UI using https://github.com/eruto-skills/marketplace, then install autopilot. To install as a standalone user skill instead:

mkdir -p ~/.agents/skills
git clone https://github.com/eruto-skills/autopilot.git ~/.agents/skills/autopilot

On Windows PowerShell:

New-Item -ItemType Directory -Force "$env:USERPROFILE/.agents/skills" | Out-Null
git clone https://github.com/eruto-skills/autopilot.git "$env:USERPROFILE/.agents/skills/autopilot"

In Codex, select the installed skill by name or invoke $autopilot with a task. In Claude Code:

/plugin marketplace add eruto-skills/marketplace
/plugin install autopilot@eruto-skills

The instructions use the tools available in the current host. Scripts are resolved from the actual skill directory, rather than a fixed author path. Additional browser, Python, or format-specific dependencies are described in SKILL.md and the references; installing the plugin alone does not install those external programs.

Maintaining the plugin package

Edit the root SKILL.md and its supporting resources, then run:

node scripts/package-plugin.mjs
node scripts/package-plugin.mjs --check

Commit the generated skills/ files with the source changes. CI checks that both layouts match, including the Claude manifest. Do not edit generated files directly.

Ver o original no GitHub

Bom para

Relacionados