convez.ai
SkillDesenvolvimento WebAvançado
happysasha18

Product Prover

happysasha18/product-prover

Revise specs como arquiteto sênior antes de codificar

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill que funciona como arquiteto sênior revisando especificações de produto através de lentes de verificação formal. Analisa PRDs, specs de features e design proposals em busca de defetos estruturais — estados incompletos, atores indefinidos, invariantes violados — e retorna achados concretos, rastreáveis à fonte, sem jargão acadêmico.

Instalação

$npx convez-ai happysasha18/product-prover

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
5
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 1 mês
Versão
v1.6.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Revisar PRDs e feature specs em busca de gaps estruturais antes da implementação
  • Validar design proposals e arquiteturas de alto nível para completude
  • Identificar edge cases, invariantes violados e fluxos de falha em especificações

Veredito

Product Prover é uma ferramenta sofisticada que traz rigor de verificação formal para revisão de especificações sem sobrecarregar com jargão — ideal para times que querem validar arquitetura antes de investimento em código. Seu rigor em rastreabilidade e concretude nas findings é raro e valioso.

README

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

product-prover

A senior-architect review of your product spec — through the lens of formal verification. A Claude Code skill.

Point it at a PRD, feature spec, HLD, LLD, or design proposal, and it reviews the document the way a principal architect would: a short opening assessment, the structural model it extracted, the gaps that matter most, and what to fix before you build.

It thinks in formal-verification primitives — entities, states, transitions, invariants, safety, liveness, atomicity, composition — but it never lectures you in jargon. The framework stays private; what you get back is in operational terms you can act on.


The rule it won't break

Never produce a finding the reader can't trace back to the document.

Every finding quotes the source and pins its location. Every consequence is concrete — who is affected, what triggers the failure, what goes wrong, what they see — not "this could be a problem." Every proposed fix is a specific artifact or decision; vague verbs (define, ensure, handle, consider) are banned. When the doc is too vague to support a concrete consequence, it says so plainly — "the spec needs to state X" — instead of inventing one.

Same instinct as its sibling track-coach: facts over plausible fiction, and the decision always stays with the author.


What it does

A continuous, structured pass — no pausing between phases:

  • Triage — is this even an analyzable spec, or marketing copy? Says so up front.
  • Opening assessment — the 30-second verdict: what it's trying to do, what's working, what needs attention, how close it is to buildable.
  • The model — extracts entities, states + transitions, actors, and composition boundaries. Tells you exactly what it assumed where the doc was ambiguous.
  • Structural issues — incomplete state space, undefined actors, components mixing roles, abstraction problems.
  • Property analysis — safety (invariants, pre/postconditions, atomicity, rollback), liveness (dead-ends, termination, silent failure masking), enforceability, and internal consistency — plus generative stress-testing against six families of questions, not pattern-matching.
  • Acknowledged gaps — the Open Items and TBDs the doc already flags, kept separate so you see what you missed first.
  • Human + operational factors — observability, cognitive load, debuggability.
  • Closing summary — top 3 things to fix, properties to state explicitly (phrased so you can paste them straight in), and the genuine open questions only you can answer.

Findings come tagged severity · plain-label (formal-term) — must-fix, should-clarify, or worth-considering — with severity reflecting real production impact, not formal imperfection.


What's inside

No code, no dependencies, nothing to build or install — product-prover is a single SKILL.md: a structured set of review instructions Claude follows. Drop it in, point Claude at a spec, read the findings. Works anywhere Claude Code runs.


Usage

This is a Claude Code skill — drop the folder into ~/.claude/skills/product-prover/ and just ask:

"review this spec" · "poke holes in this design" · "is this PRD ready / what did I miss?" · "Product Prover this"

It also has a glossary mode — /glossary liveness, /define atomicity, or just "what does composition mean?" — for the formal terms, each with a plain definition, an example, and the question it prompts in a review.


License

MIT © Alexander Abramovich.

Ver o original no GitHub

Bom para

Relacionados