convez.ai
SkillProdutividadeIntermediário
pt-hieu

Skills de Engenharia - Claude Code

pt-hieu/skills

Rails de qualidade para fluxo Claude Code: planos auditados, debug rigoroso, prompts confiáveis

Ver no GitHub
InicialPluginScore · de 0 a 100

Conjunto de skills para Claude Code que implementa rigor em fluxo de engenharia, comunicação e design. Inclui auditoria de planos com subagentes, debugging sistemático, técnicas de prompting confiáveis, commits estruturados e voice de comunicação outbound.

Instalação

$npx convez-ai pt-hieu/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
1
Forks
0
Acompanhando
1
Saúde
28%
Último commit
há 3 dias
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Auditar planos técnicos antes de implementação com análise de arquitetura e causa raiz
  • Debugging sistemático com frame do problema e validação de hipóteses
  • Design de prompts e agents com técnicas de calibração de confiança e prevenção de alucinação
  • Commits estruturados que bloqueiam batch/summary commits
  • Comunicação outbound (Slack, PRs, Jira) com voz técnica consistente
  • Documentar princípios de UI design para reutilização em projetos

Veredito

Skills pragmáticas que endereçam gargalos reais em fluxos LLM: auditoria de planos via subagentes, debugging metodizado, e voz de comunicação. Bem documentado com exemplos de workflow. Depende bastante de padrões pessoais do autor, mas os conceitos (challenge antes de commit, diagnose estruturado, prompting calibrado) têm valor geral para engenheiros que usam Claude intensivamente.

README

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

skills

Brian Pham's personal Claude Code skills, published as a plugin marketplace.

Install

Add as a marketplace in Claude Code:

/plugin marketplace add pt-hieu/skills
/plugin install brian@brian-skills
/plugin install voice@brian-skills
/plugin install design@brian-skills

How I use these — my workflow

Most of my work happens through Claude Code, where Claude does the typing and I steer. These skills are the rails that keep that loop honest. They're organized around the moments where I most often go off the road.

1. Before I let Claude commit to a plan — challenge + pitch

The most expensive mistake in an LLM session is a plausible plan that's wrong. Once it's in context, the model justifies it instead of questioning it. So before I exit plan mode, two things happen:

  • challenge runs two fresh-context subagents — one for architecture fit, one for root cause — against the plan. Two green passes is a high bar; the first plan rarely clears it, which is the point.
  • pitch then summarizes the surviving plan back to me in everyday language: problem, fix, new behavior, no jargon. That's where I catch direction-level mistakes that a technical plan would have hidden behind file names.

Only after both do I approve and let implementation start.

2. While Claude is debugging — diagnose

When something breaks, Claude defaults to the first plausible cause. diagnose is a script it follows instead: frame the problem, deepen past the first hypothesis, identify the defect class, validate with a test, run a devil's-advocate pass. It pairs with challenge — diagnose builds the hypothesis, challenge stress-tests it.

3. When Claude is writing prompts or agents — prompting

I build a lot of LLM agents. prompting is the set of techniques I keep reaching for: conflict detection, confidence calibration, pro/con balance, chain-of-verification, citation, hallucination prevention. Useful when designing agent prompts and structured-output schemas, not only when running them.

4. When the work is done — commit

Stops Claude from writing the kind of commit message a future git blame won't thank me for. Enforces format, blocks batch/summary commits, splits when concerns differ.

5. When I'm posting outward — voice

Slack messages, PR descriptions, Bitbucket and Jira comments. voice writes them in my warm, hedged, emoji-aware engineering register instead of the flat tone an LLM defaults to. Strictly outbound — not for code, plans, or chat replies.

6. When I'm building UI — design

principles collects the UI design rules I otherwise re-explain on every project: component patterns, formatting, interaction defaults. Loads automatically when Claude is touching frontend.

What's inside

Three plugins, each a coherent domain:

brian — engineering workflow rigor

SkillPurpose
challengeAudit a plan or implementation with two independent opus subagents.
commitStructured commit workflow.
diagnoseSystematic root-cause debugging methodology.
promptingResearch-backed prompting techniques for reliable LLM agents.

voice — outbound communication

SkillPurpose
voiceTeam-facing writing voice for Slack, PRs, Bitbucket, Jira.
pitchPlain-language plan pitch before exiting plan mode.

design — UI design rules

SkillPurpose
principlesUI design principles, component patterns, formatting, interaction rules.

Layout

.claude-plugin/
  marketplace.json
plugins/
  brian/
    .claude-plugin/plugin.json
    skills/
      challenge/
      commit/
      diagnose/
      prompting/
  voice/
    .claude-plugin/plugin.json
    skills/
      voice/
      pitch/
  design/
    .claude-plugin/plugin.json
    skills/
      principles/

License

MIT

Ver o original no GitHub

Bom para

Relacionados