Agent Skills para Claude Code e CLI
jerseycheese/agent-skills
Workflows de dev-loop compartilhados: testes, reviews, CI-fix e ship automatizado
Ver no GitHubColeção 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-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 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
bdgCLI. 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.comURLs. - 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/skillspaths. - 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.
| Tool | User-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.