convez.ai
SkillDesenvolvimento WebIntermediário
andreiRS

Skills — Coleção de Competências para Claude

andreiRS/skills

16 skills modulares para planejamento, spec, arquitetura, TDD e testes

Ver no GitHub
InicialPluginScore · de 0 a 100

Conjunto de 16 skills reutilizáveis em TypeScript para estender capacidades do Claude e outras LLMs. Cobre todo ciclo de desenvolvimento: planejamento, especificação, arquitetura, implementação TDD, testes e documentação. Cada skill é uma instrução concisa e composável.

Instalação

$npx convez-ai andreiRS/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
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 1 mês
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Conduzir entrevistas estruturadas e validar planos com perguntas críticas
  • Gerar especificações enxutas e arquivos de arquitetura (ADRs)
  • Implementar funcionalidades com TDD e orquestrar múltiplos agentes
  • Testar aplicações em produção e executar testes exploratórios adversariais
  • Quebrar mudanças em commits atômicos e gerar documentação HTML

Veredito

Excelente coleção de prompts de alta qualidade e bem testados para engenheiros que querem automação de processos de desenvolvimento completos. Abordagem opinionada mas flexível; valor real em orquestração TDD e análise arquitetônica.

README

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

Skills

A personal collection of reusable skills for Claude Code and other LLM tools. The source for the banner above lives in assets/skills-map.html.

Skills

SkillDescription
interviewHigh-level interview to align on the shape of a plan (~3-6 questions)
poke-holesFocused review that hunts for weak spots and hidden assumptions (~6-12 questions)
grill-meExhaustive walk of every branch of the decision tree, hands off to to-spec
to-specCapture the why and what into a lean six-section spec under docs/specs/ (capture-first)
domain-docsMaintain the GLOSSARY.md glossary and docs/adr/ decision records
improve-architectureFind deepening opportunities (shallow→deep modules) and present them as a visual HTML report
to-htmlRender what's been discussed as a single self-contained HTML page — explainer, matrix, dashboard, cheat-sheet
to-issuesBreak a plan or spec into independently-grabbable issues as thin vertical slices
implement-tddBuild an existing plan test-first, one behavior at a time, red-green-refactor
orchestrate-tddDrive a list of issues to completion by spawning one TDD agent per slice, sizing the model, verifying and reviewing each result
prove-itDrive the running app like a human to confirm the changed flows work end to end (pass/fail verdict)
break-itAdversarial exploratory testing — drive the running app to find the bugs a human QA would catch
handoffSummarise the current conversation into a handoff doc for the next session
to-atomic-commitsSplit uncommitted changes into logical atomic commits, respecting detected commit conventions
garminDownload Garmin Connect activity splits and summarize workouts
brand-guidelinesApply Andrei's personal brand colors, typography, and visual style to any rendered artifact

Size at a glance

Each skill's SKILL.md — the instruction prose only — measured to keep us honest about conciseness. Tokens are estimated as bytes ÷ 4. Regenerate after any change with bun scripts/skill-sizes.ts --write.

SkillLinesWords~Tokens
orchestrate-tdd1461,7822,729
implement-tdd1501,3441,965
show-it601,0321,561
to-html629611,483
improve-architecture698891,428
domain-docs1058721,409
to-spec1058781,401
garmin767611,268
to-issues1027281,132
to-atomic-commits81591937
brand-guidelines67396679
break-it37386619
grill-me31359567
poke-holes37321503
prove-it29328493
interview27228367
handoff24173257
Total1,20812,02918,798

Requirements

Most skills are prose-only and need nothing beyond Claude. These few shell out to external tools — install them before use:

ToolUsed byInstall
bungarmin, show-it, installing agent-browserbun.sh
agent-browsergarmin, prove-it, break-itbun add -g agent-browser && agent-browser install
Google Chrome (real app)garmingoogle.com/chrome

garmin drives real Chrome with a persistent authenticated profile (to pass Garmin's Cloudflare bot check) and always runs headed — see its SKILL.md for the one-time sign-in.

From idea to shipped code

The plan-and-build skills chain together. Pick the review depth that fits, turn the result into work, then implement it:

  1. Shape the plan — three depth tiers for stress-testing a plan, pick based on how deep you want to go:
    • interview → quick alignment, big-picture only
    • poke-holes → probe the 2-3 riskiest parts
    • grill-me → exhaustive, then hands off to to-spec
  2. Capture it — to-spec structures the resolved plan into a lean six-section spec under docs/specs/. Capture-first: it records what's been decided, asking only for missing required pieces.
  3. Break it down — to-issues slices the spec into independently-grabbable issues.
  4. Build it — implement-tdd consumes the spec or issues and builds them test-first.
  5. Check it in the real app — once it's built and the suite is green, drive the running app like a human: prove-it confirms the changed flows work end to end; break-it throws nasty inputs and edge cases at them to find what a human QA would catch. Distinct from the test suite (implement-tdd/orchestrate-tdd) — this is the manual, browser-driven pass.

For a batch of issues you'd rather not build by hand, orchestrate-tdd sits one level above implement-tdd: instead of coding, it spawns one agent per slice (running implement-tdd on that slice), sizes each to the cheapest model that fits, then independently re-runs the tests and reviews the diff before advancing. Reviews are tier-gated (skip trivial, simplify for simple, code-review for complex, plus a final cross-slice pass), failures retry then escalate to the most-capable model, and it writes status back to the tracker as each slice lands. Use it to drive a whole to-issues backlog to a reviewed branch autonomously.

domain-docs runs alongside this flow as the keeper of durable domain knowledge: the GLOSSARY.md glossary and docs/adr/ decision records. grill-me and to-spec flag fuzzy terms and hard-to-reverse decisions and point here to record them; it also works standalone.

improve-architecture feeds the front of this flow for existing codebases: it surveys the code, surfaces the strongest deepening opportunities as a visual HTML report, and stops. Take a chosen candidate into grill-me to design it, or to-issues to slice it.

handoff supports this flow across sessions: when work spans multiple sessions or collaborators, write a handoff doc so a fresh session can pick up where the last one left off.

Other

  • garmin → pull recent Garmin Connect activity data and summarize it.

Inspiration

This collection draws on the ideas, structure, and conventions of:

Ver o original no GitHub

Bom para

Relacionados