Vibe Coding Skills
drakep/vibe-coding-skills
Skills para os problemas "after it works" que AI não sabe questionar
Ver no GitHubConjunto de skills drop-in para Claude Code que cobrem os aspectos críticos que IA geralmente ignora ao gerar código — segurança, custo e confiabilidade. Inclui auditorias de segurança básicas que identificam chaves expostas, bancos de dados abertos, falta de rate limiting e outros problemas que viralizam depois que o projeto vai para produção.
Instalação
npx convez-ai drakep/vibe-coding-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
- Auditar projetos vibe-coded antes de deployar em produção
- Validar segurança de APIs, autenticação e permissões de banco de dados
- Identificar exposição de chaves e secrets em frontend e git history
- Implementar rate limiting e proteções contra abuso de custos
Veredito

Aborda honestamente o gap entre "código que funciona" e "código pronto para produção". Essencial para quem constrói com IA — mais prático que teórico, com checklist executável mesmo sem agent.
README
A tradução para o português deste README ainda está na fila.
Vibe Coding Skills
A collection of skills for people who build with AI.
Vibe coding is real now. You can describe an app and have it shipped in an afternoon. That part is solved. The part that isn't solved is everything that happens after "it works on my screen" — the boring, unglamorous stuff that decides whether your project quietly runs for years or gets wrecked the first week it gets traffic.
These are the skills the AI usually skips because you didn't know to ask for them. Each one is a drop-in Claude Code skill — point your agent at the skills/ folder and it can run them on your project on demand.
Why every vibe coder needs these
When you build the traditional way, you pick up this stuff by osmosis — code reviews, getting burned once, a senior dev yelling at you in a PR. When you vibe code, none of that happens. The model writes confident, clean-looking code that runs, and you ship it. It looks done. It usually isn't.
The gap is never "does it work." It's:
- Security — the model will happily put your API key in the frontend, leave your database wide open, and skip rate limiting, because you asked for a feature, not for it to be safe.
- Cost — one exposed key or one unthrottled endpoint and you wake up to a four-figure bill.
- Trust — the moment a real user (or a bored teenager with Burp Suite) pokes at it, the cracks show.
None of this requires you to become a security engineer. It requires a checklist and ten minutes. That's what these skills are.
Skills
| Skill | What it does |
|---|---|
basic-security-audit | Walks your project and flags the things that actually get vibe-coded sites hacked — exposed keys, open databases, missing rate limits, broken auth, and more. Severity-ranked, with fixes. |
More coming. Backups, deploy hygiene, cost guardrails — the rest of the "after it works" list.
How to use
If you use Claude Code, drop this repo's skills/ folder into your project (or your ~/.claude/skills/) and ask:
run a basic security audit on this project
If you don't use an agent, the skill files are just well-structured Markdown checklists. Open skills/basic-security-audit/SKILL.md and work down it by hand. Everything in there is a thing you can check yourself in a browser, a terminal, or your hosting dashboard.
The one-minute version
If you read nothing else, check these four before you ship anything:
- No secret keys in your frontend. Open your deployed site, hit
View Source/Inspect, search the page and JS bundles forsk_,key,secret,firebase,supabase. If a secret key shows up, anyone can use it. (Public/anon keys are fine — see #2.) - Your database isn't wide open. If you use Supabase, Row Level Security must be on. If you use Firebase, your rules must not be
allow read, write: if true. The public key is supposed to be public — the rules are what keep people out. This is the single most common way vibe-coded apps get owned. - Rate limiting exists. Your signup and contact forms should cap requests per IP. Without it, one script turns into thousands of fake accounts or emails, and your bill or your inbox pays for it.
- No secrets committed to git.
git log -p | grep -iE 'key|secret|token|password'. If your.envis in your history, the keys in it are burned — rotate them, even if the repo is private.
That's 80% of the risk in four checks. The full audit catches the rest.
Built by people who ship fast and would prefer not to get hacked. PRs welcome — if you've seen a vibe-coded mistake in the wild that isn't covered, add it.