PM Skill — Orquestrador de Projetos para Claude
bendusz/pm-skill
PM orchestrador que gerencia projetos via suagentes especializados
Ver no GitHubTransforma Claude em um gerente de projeto disciplinado que descobre requisitos, especifica, planeja, obtém aprovação e orquestra a construção através de suagentes especializados — sem escrever o código. Segue um fluxo repetível: descoberta → especificação → clarificação → planejamento → aprovação → decomposição → implementação → revisão → verificação → entrega.
Instalação
npx convez-ai bendusz/pm-skill1 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
- Gerenciar projetos de software do zero com descoberta e planejamento estruturado
- Orquestrar desenvolvimento paralelo de histórias em git worktrees independentes
- Garantir rastreabilidade entre requisitos, plano e implementação com análise de cobertura
Veredito

Skill robusta para quem precisa rigor metodológico e delegação confiável em projetos complexos. Cria artefatos duráveis (specs, planos, logs) e mantém o PM como orquestrador puro — ideal para brownfield e equipes distribuídas, mas exige disciplina no setup inicial.
README
A tradução para o português deste README ainda está na fila.
pm-skill — a Project/Product Manager skill for Claude Code
Turn Claude into a disciplined Project / Product Manager that discovers, specifies, plans, gets your sign-off, decomposes work into stories, and orchestrates the build through specialist subagents — gate, review, fix, verify, ship, and log — without writing the code itself.
One repeatable way of working:
discover → specify → clarify → plan → sign-off → analyze → decompose → build → gate → review → verify → ship → log
It is generic and self-contained: it works on a bare Claude Code install and gets richer if you happen to have other tools.
Install
Run these two steps separately — submit the first, wait for it to finish, then run the second (don't paste both at once, or the second line gets swallowed into the first command's argument):
1. Add the marketplace
/plugin marketplace add https://github.com/bendusz/pm-skill
2. Install the plugin
/plugin install pm-skill@pm-skill
Use the full https:// URL above — the owner/repo shorthand resolves to SSH, which fails on
machines without a GitHub SSH key/host key set up.
If it doesn't appear right away, restart your Claude Code session.
Use
- Just describe the work — e.g. "act as my PM to build a CLI todo app" — and the
project-managerskill activates, or - run the command explicitly:
/pm-skill:pm build a CLI todo app
The PM runs discovery with you, writes a plan, and waits for your explicit sign-off before building anything.
How it works
| Phase | What happens |
|---|---|
| Discovery | You and the PM agree the problem and the best solution. |
| Specification | A durable docs/spec.md — user stories, requirements, acceptance criteria, success metrics (what & why, not how). |
| Clarification | Open [NEEDS CLARIFICATION] questions resolved one at a time before planning. |
| Plan & sign-off | A written docs/plan.md that derives from the spec (with traceability); you approve it before any code. |
| Analyze | A read-only cross-artifact consistency check (coverage, contradictions, constitution) after the plan, before decomposition — optionally before sign-off. |
| Decomposition | Sprints → self-contained story files under docs/stories/, each tracing to requirement IDs. |
| Implementation loop | Per story: build → gate → review → fix → verify → ship → log, run by subagents. |
| Parallel stories | Independent [P] stories can build at once in isolated git worktrees, then integrate one at a time (opt-in; safe fallback to sequential). |
| Review & verification | A separate read-only reviewer + the project's real test/lint/build gates + a final read-only pm-verifier PASS; bounded fix loops. |
| Logging | A tmp/log.md logbook + tmp/pm-state.json so a lost session can resume. |
Bundled specialist agents do the work — a builder (expert-builder), a risk-selected read-only
review panel (code-integrity-reviewer, architecture-reviewer, security-auditor),
a test-engineer (tests only), a debugger (read-only root-cause → fix plan), a read-only
final pm-verifier (independent PASS/FAIL before ship), a technical-writer (docs only), and
a codebase-analyst for brownfield work. The PM stays an orchestrator and protects its own
context by handing each agent only what it needs.
Default check-in is sprint-level (you review at each sprint boundary); configurable to
story-level or fully autonomous. Pick a scale (tiny→regulated) to right-size the workflow —
tiny work stays lightweight; regulated work makes every gate mandatory.
Commands
| Command | What it does |
|---|---|
/pm-skill:pm | Act as the PM end to end — discover, plan, get sign-off, orchestrate delivery. |
/pm-skill:specify | Capture/refine docs/spec.md — the product spec (what & why). |
/pm-skill:clarify | Resolve open [NEEDS CLARIFICATION] in the spec, one question at a time (≤5). |
/pm-skill:constitution | Create/update docs/constitution.md — project-specific governing rules. |
/pm-skill:analyze | Read-only consistency & quality report across all artifacts (never edits). |
/pm-skill:checklist | Generate/evaluate a spec/plan/story/verification quality checklist under docs/checklists/. |
/pm-skill:doctor | Check environment readiness (toolchain, deps, gates run) before building. |
/pm-skill:resume | Read saved state + logbook and continue where you left off. |
Artifacts
Committed under docs/ (authoritative):
docs/spec.md— product specification (user stories, requirements, acceptance criteria, metrics).docs/plan.md— delivery plan, derived from the spec with traceability.docs/stories/*.md— self-contained story files, each tracing to requirement IDs.docs/constitution.md— project-specific governing principles (optional).docs/checklists/*.md— spec/plan/story/verification quality checklists (optional).docs/verification/*.md— per-story verification reports (optional; recommended for non-trivial work).
Runtime-only under tmp/ (gitignored, disposable):
tmp/log.md— the logbook.tmp/pm-state.json— machine-readable state for resume.tmp/environment-check.md—/pm-skill:doctor's readiness report.
Safety
- No implementation before your sign-off.
- Repository safety: the PM never overwrites your files without asking, commits only what it
created for the current story, runs
git initonly after asking, and never pushes without an explicit request. - Optional hardening: for mechanical enforcement (a read-only Bash posture, sign-off), the bundled hardening guide uses plain Claude Code permissions/hooks — no external process or dependency.
Optional enhancements (work alongside — not required)
pm-skill is fully functional on its own. If your environment also has any of these, the PM may prefer them where useful — but nothing here is a dependency:
- A dedicated planning / TDD skill suite for richer discovery and planning.
- An external code-review tool (for example an OpenAI Codex–based reviewer, or another model's CLI) for the optional independent review step.
ghplus a GitHub remote for real pull requests (otherwise the PM uses local merges).
The spec / clarify / analyze / constitution steps add spec-driven rigor (inspired by spec-driven
development tools), and pm-verifier adds an independent final check — but they are built in and
self-contained. pm-skill does not depend on spec-kit, and there is no external verifier
harness and no separate process in this workflow: pm-verifier is an ordinary read-only Claude
Code subagent the PM dispatches.
License
MIT — see LICENSE.