Parthiv Skills — Guardrails de Engenharia para Claude Code
parthiv9817/parthiv-skills
Guardrails comportamentais para Code: verifica, escopeia, delega com rigor.
Ver no GitHubConjunto portátil de guardrails comportamentais para Claude Code, destilado de dois meses de erros documentados pelo autor e validado contra guidelines de Andrej Karpathy e orientações Anthropic. Oferece 5 skills modulares (verificação antes de afirmar, economia de investigação, escopo cirúrgico, delegação de modelos, processo colaborativo) que funcionam como camada de disciplina sob qualquer metodologia.
Instalação
npx convez-ai parthiv9817/parthiv-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
- Evitar afirmações precipitadas sobre root causes ou status de conclusão
- Sair de loops custo-benefício ruins de guess-fix-retry
- Manter escopo cirúrgico ao editar código compartilhado ou config
- Rotear corretamente entre modelos ao spawnar subagents
- Coordenar fallbacks, multi-tentativas e decisões de 'bom o suficiente' em equipe
Veredito

Skill maduro e prático, baseado em falhas reais pagadas pelo autor e validado contra autoridades técnicas. Funciona como camada comportamental sob metodologias pesadas, com instalação simples via symlink.
README
A tradução para o português deste README ainda está na fila.
parthiv-skills
A portable, project-agnostic set of engineering guardrails for Claude Code — distilled from two months of my own documented mistakes, cross-checked against Andrej Karpathy's coding guidelines and Anthropic's subagent/cost guidance.
The point: the best skills to encode are the ones you've already paid for in
real failures. These ride with me into every project, alongside superpowers.
Architecture (hybrid)
Two layers, by token economics:
- Always-on core — the ~8 highest-frequency, universal one-liners live in my
global
~/.claude/CLAUDE.md(loaded every turn; kept terse on purpose). - Modular on-demand skills — the detailed versions live here as separate
SKILL.mdfiles, loaded only when the situation calls for one (progressive disclosure, likesuperpowers).
The skills
| Skill | Use when |
|---|---|
verifying-before-claiming | About to state a root cause, declare "works/done", or read ambiguous evidence |
investigation-economics | Stuck in a guess-fix-retry loop on a costly cycle; about to ask an AI before searching |
surgical-scope | Over-building for the stage; editing config/shared state; touching code you didn't write |
delegation-and-model-routing | Spawning subagents / fanning out; choosing which model a worker runs on |
process-and-collaboration | Fallback paths, multi-attempt debugging, committing, risky ops, "good enough" calls |
Install
These are personal Claude Code skills. To make Claude Code load them, symlink (or copy) each into your personal skills directory:
mkdir -p ~/.claude/skills
for d in ~/Documents/parthiv-skills/skills/*/; do
ln -sfn "$d" ~/.claude/skills/"$(basename "$d")"
done
(Symlinks let you keep editing in the repo and push to GitHub while Claude Code loads the live versions.)
Relationship to superpowers
Complementary, not competing. superpowers provides heavy process methodologies
(TDD, systematic-debugging, writing-plans). parthiv-skills provides behavioral
guardrails — the verify/scope/delegation discipline that sits underneath any
methodology. Run both.
Provenance
- DS1 — Karpathy's
andrej-karpathy-skills(Think-Before-Coding, Simplicity-First, Surgical-Changes, Goal-Driven-Verification) + Anthropic subagent/model-routing guidance. - DS2 — my own
diaries/+mistakes/+epilogue/corpus (the failures these guardrails are built to prevent).
The two sources converged on the same principles independently — which is the reason to trust them.