Claude Skills — Workflows de Código Extensíveis
payne0420/claude-skills
Skills engine-agnostic para workflows, delegation e auditoria de agentes Claude
Ver no GitHubColeção de skills reutilizáveis para Claude Code que implementam workflows de desenvolvimento (scope-to-ship, build loops, evals) e delegação a agentes externos (Codex, Cursor, opencode). Oferece engine-agnostic delegation, permitindo que Claude escolha o melhor agente por tarefa, além de ferramentas para auditoria de loops autônomos, tracking de uso e acesso a docs atualizadas.
Instalação
npx convez-ai payne0420/claude-skills1 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
- Automatizar fluxo de feature completo (scope → design → ship) com um único agente alinhado
- Delegar trabalho a agentes externos (Codex, Cursor, opencode) mantendo isolamento
- Desenhar e auditar loops autônomos com gates de qualidade e state management
- Evaluar outputs de IA contra rubrics e converter falhas em testes permanentes
- Rastrear uso e limites de rate across múltiplos harnesses de código
Veredito

Conjunto maduro e bem-arquitetado de skills para elevar a automação com Claude Code além do trivial. O design engine-agnostic é robusto; requer familiaridade com CLI tooling e workflows de CI/CD para valor máximo.
README
A tradução para o português deste README ainda está na fila.
claude-skills
A collection of Claude Code skills — reusable,
model-driven workflows that extend the agent. Each folder is a self-contained skill
(a SKILL.md plus any templates/scripts) that Claude Code loads on demand.
A guiding principle across these: engine-agnostic. Where a skill can delegate work to another coding agent, it offers a neutral menu (Codex / Cursor / opencode / a raw endpoint) and lets Claude pick per task — nothing is hardcoded to one vendor.
Skills
Build workflows
| Skill | What it does |
|---|---|
scope-to-ship | Ship a feature with one aligned agent: discuss + scope → write a tiny ADR → execute against a Master PRD → dogfood end-to-end. The anti-fan-out methodology. |
ultracode-external-agents | Author ultracode workflows whose agent() steps run on an external CLI (Codex / Cursor / opencode) instead of Claude subagents — bridge any step onto another model. |
loopify | Design, scaffold, and audit autonomous agent loops: qualify a recurring task (and refuse when it doesn't fit), build the minimum viable loop (automation + standing context + state file + objective gate), dry-run, then schedule. |
Delegate to another coding agent
| Skill | What it does |
|---|---|
codex-exec | Drive OpenAI's Codex CLI headlessly (codex exec) — independent implementer, reviewer, or second opinion behind a hard OS sandbox. |
cursor-agent | Drive Cursor's CLI agent headlessly (cursor-agent -p) — fast repo-aware edits/reviews, with git-worktree isolation. |
opencode | Drive the opencode CLI (opencode run) — a model-agnostic second agent; pin any provider/model. |
llm-endpoint | Call a configured OpenAI-/Anthropic-compatible HTTP endpoint as a lightweight model backend (a raw completion, not an agent). |
Quality & docs
| Skill | What it does |
|---|---|
eval-loop | Stand up an eval loop: score AI output against a rubric, gate what's below the line, turn every failure into a permanent test. |
find-docs | Fetch current library / framework / SDK / CLI docs via Context7 — instead of trusting training-data API surface. |
Usage tracking
| Skill | What it does |
|---|---|
harness-usage | Read usage / rate-limits / cost across other coding harnesses (Codex, Cursor, opencode, Gemini, …) via CodexBar. |
track-usage | Track this Claude Code session's own 5-hour / weekly / context usage and when each resets. |
Install
Skills live in a skills/ directory that Claude Code scans — either your personal
~/.claude/skills/ or a project's .claude/skills/. Clone this repo and link (or
copy) the skills you want into place.
git clone https://github.com/payne0420/claude-skills ~/Github/claude-skills
# enable one skill for yourself (symlink so `git pull` keeps it current):
ln -s ~/Github/claude-skills/scope-to-ship ~/.claude/skills/scope-to-ship
# …or enable all of them:
for d in ~/Github/claude-skills/*/; do
[ -f "$d/SKILL.md" ] && ln -sfn "$d" ~/.claude/skills/"$(basename "$d")"
done
Then invoke a skill in Claude Code by name (e.g. /scope-to-ship) or just describe
what you want — Claude routes to the matching skill via its description.
Notes
- Some skills wrap external CLIs (
codex,cursor-agent,opencode,ctx7,codexbar). Those tools must be installed and authenticated separately; each skill checks for its dependency and tells you if it's missing. llm-endpointreads its base URL + API key from a local file (~/.config/llm-endpoint/env) that is never committed — seellm-endpoint/config.example.sh.
License
MIT.