stavxyz/skills — Validação e Polimento de PRs
stavxyz/skills
Validação de specs e polimento automático de PRs com revisão dupla
Ver no GitHubConjunto de skills para Claude Code que automatiza revisão de especificações e pull requests. Inclui `/validate` para auditar specs contra a base de código (fact-check + SOLID/hygiene) e `/polish-pr` para revisar, corrigir e mergear PRs de forma autônoma, rodando dois revisores em paralelo.
Instalação
npx convez-ai stavxyz/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
- Validar especificações técnicas contra código existente antes de implementação
- Revisar e corrigir automaticamente pull requests antes do merge
- Auditar design para violações SOLID e problemas de higiene de código
- Acelerar ciclo de code review com múltiplos revisores paralelos
Veredito

Skill robusto para equipes que usam Claude Code, com lógica sofisticada de triagem e gates de aprovação. Exige familiaridade com Claude Code e fluxos Git, mas elimina gargalos manuais em validação de design e revisão de PRs.
README
A tradução para o português deste README ainda está na fila.
stavxyz/skills
A small Claude Code plugin marketplace with two multi-reviewer skills:
| Skill | Command | What it does |
|---|---|---|
| validate | /stavxyz:validate <path-to-spec-or-plan.md> | Runs two reviewers in parallel against your codebase — a fact-check pass (are the spec's claims about existing code true?) and a SOLID / hygiene pass (is the design direction sound?) — then addresses the findings in-spec, with gates for deferrals, Critical corrections, and net-negative design. |
| polish-pr | /stavxyz:polish-pr <PR#> | Rebases a PR, runs two independent code reviews in parallel, addresses every finding at every severity in-PR, updates docs, runs a test plan, and pushes. |
Install
/plugin marketplace add stavxyz/skills
/plugin install stavxyz@skills
The first command registers this repo as a marketplace named skills; the second installs the stavxyz plugin from it. Plugin skills are namespaced by the plugin name, so after installing they're invoked as /stavxyz:validate and /stavxyz:polish-pr.
To update later:
/plugin marketplace update skills
/plugin update
Updates aren't automatic — run those when a new version ships. Because Claude
Code caches plugins under a version-stamped path, every change to shipped skill
content (skills/) ships with a version bump; see RELEASING.md
for the why and the contributor workflow.
The skills
/stavxyz:validate
/stavxyz:validate docs/specs/2026-05-31-my-feature-design.md
Validates a spec or plan markdown file against the current HEAD of its git repo. It dispatches two general-purpose subagents in parallel using the bespoke reviewer prompts shipped alongside the skill:
- fact-check — verifies claims the spec makes about existing code (paths, symbols, behavior) against the real codebase.
- solid-hygiene — audits the design direction for SOLID/hygiene problems and flags anything net-negative (a change that would make the codebase worse).
Findings are deduped, triaged, and addressed in-place in the spec. Three conditions gate on your approval before edits proceed: deferral candidates, Critical fact-check findings, and net-negative design findings. On success the spec's frontmatter gets a validated: block recording the SHA, date, and finding counts.
On a clean bless — blessed with zero caveats (e.g. no deferrals, accepted net-negatives, skipped Critical fact-checks, hallucinated quotes, parse failures, or reviewer contradictions; both reviewers ran; and the spec/plan kind wasn't a tie-break guess) — validate auto-continues to the next stage of the pipeline: a blessed spec flows straight into superpowers:writing-plans, and a blessed plan flows straight into superpowers:subagent-driven-development. It announces this with a banner and proceeds without a prompt. If the bless carries any caveat — or the next-stage skill isn't available to invoke (not installed, or model-invocation disabled) — validate stops and recommends the next step for you to run by hand. The skill's "Auto-continue on clean bless" section holds the authoritative caveat list.
/stavxyz:polish-pr
/stavxyz:polish-pr 142
The last-mile pass before a PR merges. It rebases onto the latest upstream, runs two code reviews in parallel (from pr-review-toolkit and superpowers), and fixes every finding in-PR rather than deferring. It also updates all docs, builds a test plan into the PR description, sweeps the branch for stray AI-attribution lines, and pushes.
/stavxyz:polish-pr has prerequisites. It hard-requires both review systems to be installed and refuses to run in degraded (single-reviewer) mode:
pr-review-toolkit— provides thepr-review-toolkit:code-revieweragent /review-prskill.superpowers— providessuperpowers:requesting-code-review(5.1.0+).
Both are available from Anthropic's official plugin marketplace:
/plugin marketplace add anthropics/claude-plugins-official
/plugin install pr-review-toolkit@claude-plugins-official
/plugin install superpowers@claude-plugins-official
If either is missing, /stavxyz:polish-pr stops and tells you what to install.
Compatibility
These skills target Claude Code. Two layers are worth distinguishing:
- The packaging (
.claude-plugin/marketplace.json+/plugin install) is Claude Code's plugin format. Some other agentic CLIs are converging on compatible plugin/skill formats, but support and exact semantics vary by tool — verify against your tool's own docs before assuming portability. - The skill content uses Claude-Code-specific tool names (
Task,AskUserQuestion,Edit/Write, thegeneral-purposesubagent type) and frontmatter (disable-model-invocation). Running these under another agent would require the tool-mapping shims that ecosystems like superpowers ship.
In short: out of the box, use these with Claude Code. Other tools may need adaptation.
Manual install (without the plugin system)
If you'd rather not use the marketplace, symlink the skills into your user skills folder:
git clone https://github.com/stavxyz/skills.git
# The clone creates ./skills (the repo), and the skills themselves live under
# its skills/ subdirectory — hence the intentional skills/skills/ below.
ln -s "$PWD/skills/skills/validate" ~/.claude/skills/validate
ln -s "$PWD/skills/skills/polish-pr" ~/.claude/skills/polish-pr
Installed this way they are user skills, which are not namespaced — so you invoke them by bare name (/validate, /polish-pr) rather than the /stavxyz: prefix used for the plugin install. validate resolves its reviewer templates relative to the installed skill directory, so it works under either install method.
Recommended for development. The symlinks point at your live working tree, so editing a skill and running
/reload-plugins(or starting a new session) picks the change up instantly — no version bump, no/plugin marketplace update. Reserve the versioned plugin/marketplace flow (see RELEASING.md) for distributing to others.
Repository layout
.
├── .claude-plugin/
│ ├── marketplace.json # registers this repo as the "skills" marketplace
│ └── plugin.json # defines the "stavxyz" plugin
├── .githooks/
│ └── pre-push # blocks pushes that change skills/ without a version bump
├── RELEASING.md # versioning rule + contributor release checklist
├── skills/ # distributed to installers
│ ├── validate/
│ │ ├── SKILL.md
│ │ ├── fact-check-reviewer.md
│ │ └── solid-hygiene-reviewer.md
│ └── polish-pr/
│ └── SKILL.md
└── tests/ # dev-only, not part of the installed plugin
└── validate-fixtures/ # sample specs for exercising validate by hand
Contributing: enable the release guard once per clone with
git config core.hooksPath .githooks. See RELEASING.md.