convez.ai
SkillAutomaçãoIntermediário
jerseycheese

Agent Skills para Claude Code e CLI

jerseycheese/agent-skills

Workflows de dev-loop compartilhados: testes, reviews, CI-fix e ship automatizado

Ver no GitHub
InicialPluginScore · de 0 a 100

Coleção de skills reutilizáveis para automação de workflows em development: testes, revisão de código, triage de issues, CI/CD e limpeza de código. Funciona com Claude Code, Codex, Gemini CLI e qualquer ferramenta que leia o diretório `.agents/skills`. Cada skill é modular e invocável via frontmatter declarativo.

Instalação

$npx convez-ai jerseycheese/agent-skills

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
4
Forks
1
Acompanhando
0
Saúde
42%
Último commit
há 3 dias
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Automatizar triage e priorização de issues GitHub
  • Revisar PRs e aplicar fixes não-controversos em uma passada
  • Diagnosticar e corrigir testes falhando com limite de tentativas
  • Conduzir issue completa de análise até merge com ship-issue
  • Auditar code health, detectar dead code e simplificar tech debt
  • Validar mudanças de styling contra design tokens

Veredito

Conjunto maduro e bem-estruturado de skills para automação de dev-loop em equipes que usam Claude/Gemini CLI. Cobre desde triage até post-merge com padrões declarativos — útil para times que querem repetibilidade em workflows, mas exige disciplina para manter skills sincronizados entre plataformas.

README

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

agent-skills

A collection of agent skills usable across Claude Code, Codex, Gemini CLI, and any tool that reads a shared .agents/skills directory. Each skill lives in its own directory with a single SKILL.md whose frontmatter tells the agent when to invoke it and what to do.

Skills

  • analyze-issue — Read a GitHub issue and produce a technical spec with scope, patterns, and MVP tests.
  • browser-debugger-cli — Inspect live pages via the Chrome DevTools Protocol through the bdg CLI. Token-efficient alternative to a full page snapshot.
  • ci-fix-with-memory — Auto-invoke when CI is red. Reads handoffs and a known-issues log so the same fix isn't tried twice.
  • code-health-audit — The "embarrassing code" / overengineering sweep. Turns vague tech-debt into a short list of measurable, validated simplifications to ship.
  • cyoa — Choose Your Own Adventure mode. Frame the task as a branching story; every fork is a real design decision.
  • dead-code-cleanup — Find and remove orphaned code, stale stories, and trivial tests after verifying nothing uses them.
  • drupal-mental-model — Explain code by mapping framework concepts to Drupal equivalents, for a Drupal/theming background.
  • ds-guard — Audit styling changes against the project's design tokens and showcase. Flags hardcoded values, Tailwind utilities, and off-system tokens.
  • github-screenshot — Generate GitHub-compatible image markdown via raw.githubusercontent.com URLs.
  • kiss — Strip a diff down to the minimum surface needed for its goal. Drops the "while I'm here" extras.
  • post-merge — After a PR merges: close out linked issues, then recommend the single best next thing to work on.
  • pr-review-fix-pipeline — Review a PR and apply fixes for the non-controversial issues in one pass.
  • prioritize-issues — Rank the backlog by value, effort, age, and roadmap fit. Returns a top-5 with specs for the top 3.
  • project-bootstrap — Stamp a new repo's agent scaffolding: a CLAUDE.md from the house template, starter settings, and a launch.json with a free port.
  • ship-issue — Drive one issue from cold start to merged PR: sync, analyze, minimal fix with a test, PR, CI green, post-merge cleanup, wrap-up.
  • skill-parity — Keep skills in sync across Claude Code, Codex, Gemini CLI, and shared .agents/skills paths.
  • tdd-implement — Red-green-refactor. Tests track acceptance criteria, not implementation details.
  • test-fix — Diagnose and fix failing tests with a hard 3-attempt limit per failure to avoid debugging spirals.
  • test-workflow — Pick and run the right suite (unit, E2E, visual, CSS lint) after a source edit, before commit.
  • visual-crawl — Crawl the running app at random breakpoints, screenshot regressions, check token consistency.
  • worktree-enhanced — Set up a git worktree with branch, dev-server detection, and project-specific setup.
  • wrap-it-up — Close out a chat with a tight summary and a sweep of related context files (memory, plans, status, logs).

Layout

<skill-name>/
  SKILL.md    # frontmatter (name, description) + body (the actual instructions)

The description field is what the agent reads to decide whether to invoke the skill — keep trigger phrases concrete and current.

Install

Each agent looks for skills in a specific directory. Copy or symlink the skill folder (the whole thing, not just SKILL.md) into the right path for whichever tool you're using.

ToolUser-level (available everywhere)Project-level (this repo only)
Claude Code~/.claude/skills/<skill-name>/.claude/skills/<skill-name>/
Codex~/.agents/skills/<skill-name>/.agents/skills/<skill-name>/
Gemini CLI~/.gemini/skills/<skill-name>/ or ~/.agents/skills/<skill-name>/.gemini/skills/<skill-name>/ or .agents/skills/<skill-name>/

The .agents/skills path is the interoperable spot — Codex and Gemini both read from it, so it's the path to pick if you want one copy to cover both.

Quick install for a single skill in Claude Code:

cp -r kiss ~/.claude/skills/

Or symlink if you want updates to flow back to this repo automatically:

ln -s "$(pwd)/kiss" ~/.claude/skills/kiss

If you're juggling all three tools, the skill-parity skill handles copying, updating, and pruning across every provider path in one pass — that's the path worth taking if you're maintaining more than a couple of skills.

Claude Code via marketplace (no manual copy)

This repo is also a Claude Code plugin marketplace, so Claude can install every skill at once and pull updates with a git push instead of a manual copy:

claude plugin marketplace add jerseycheese/agent-skills
claude plugin install workflow-skills@jerseycheese-skills

After that, claude plugin marketplace update (or a restart) picks up new commits. Codex and Gemini still use the copy/symlink paths above; this repo stays the single source for all of them.

Installed skills are namespaced as workflow-skills:<name>. They're model-invoked — Claude runs them automatically when your request matches the skill's description, so they won't show up in the / slash-command menu. Browse them with /skills, or just describe the task (e.g. "prioritize my issues").

After installing, restart the agent (or run /skills reload in Gemini) so the new skill is picked up.

Ver o original no GitHub

Bom para

Relacionados