convez.ai
SkillAutomaçãoAvançado
musharna

Claude Skills para Agentes de Código

musharna/claude-skills

Skills de disciplina para agentes IA evitarem erros sistemáticos em código

Ver no GitHub
BomPluginScore · de 0 a 100

Três skills de rigor para codificação agêntica no Claude Code que reforçam disciplina de raciocínio. Combatem falhas típicas de agentes IA: pular para correções sem diagnosticar a causa, afirmar fatos sem fundamentação e desperdiçar tokens em screenshots brutos.

Instalação

$npx convez-ai musharna/claude-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
0
Saúde
85%
Último commit
há 3 semanas
Versão
v0.1.0
Issues abertas
1
Contribuidores
1

Casos de uso

  • Diagnosticar bugs com causalidade estruturada (causal-vs-bandaid)
  • Validar citações e fatos em documentos técnicos com fundamentação (grounded-quote-or-decline)
  • Otimizar tokens em automação com browser explorando árvore de acessibilidade (playwright-token-budget)

Veredito

Ferramenta especializada para quem constrói agentes IA no Claude Code. Força disciplina de raciocínio e impede desperdício de contexto — altamente relevante para pipelines de codificação automática, mas nicho.

README

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

claude-skills

Three general-purpose rigor skills for agentic coding in Claude Code — packaged as an installable plugin.

validate License: MIT

These are reasoning-discipline skills, not domain tools. Each one counters a specific default failure mode of an LLM agent: jumping to a fix before finding the cause, asserting facts it can't ground, and burning tokens on raw browser screenshots.

Install

/plugin marketplace add musharna/claude-skills
/plugin install claude-skills@musharna-claude-skills

Claude Code will load the three skills and invoke them automatically when their trigger conditions match (or you can invoke one by name).

The skills

causal-vs-bandaid

Fires before you state a root cause or propose a fix. Forces a 3-hypothesis gate (enumerate three candidate mechanisms before naming one), a one-sentence causal chain, and a "mechanism test" that distinguishes a real fix from a tripwire removal that leaves the bug's mechanism intact. Counters the pull toward the smallest-scope patch that makes the test pass.

Symptom: requests intermittently 500 after a deploy. H1 (cache): stale config cached past TTL. For: timing lines up with restarts. Against: errors persist past one TTL. H2 (migration): new column nullable, code assumes non-null. For: stack trace hits that field. Against: only some rows. H3 (connection pool): pool sized below worker count, exhaustion under load. For: errors cluster at peak. Against: pool metrics look idle. Going with H2 — discriminator: failing requests all touch rows written before the migration. Fix corrects the column default at the schema layer, not a null-guard at the call site.

grounded-quote-or-decline

Fires when you're about to assert a sourced fact in load-bearing prose (audit memos, RFCs, wiki pages, citations). Every substantive claim either ships with a verbatim quote tied to a locatable source, or is declined — there is no paraphrase-without-quote third state. Pairs with the ghostcite CLI for deterministic citation byline checks.

Grounded: "The default request timeout is 30s." → quote: "DEFAULT_TIMEOUT = 30 # seconds" — pointer: src/client.py:12. Declined: "The library was benchmarked at 40% faster than the alternative." → no source located; rewritten as "Relative performance vs the alternative is unverified" (or cut).

playwright-token-budget

Fires before driving the Playwright MCP. Encodes the five quantitative DOM-serializer rules from browser-use that cut a typical page from ~150k tokens to ~1.5k — default to the accessibility tree, not screenshots; drop non-content and decorative nodes; select interactive elements only.

Task: click the "Checkout" button. Tempting: browser_take_screenshot (full page) to "see" the layout → ~30k tokens, opaque. Routed: browser_snapshot → a11y tree lists button "Checkout" [ref=e42] in ~2k tokens → click e42 directly. Screenshot only if the deliverable is the picture itself, and then with fullPage: false + a narrow element selector.

Companion

musharna/fresh-read-guard — hooks that enforce read-before-write, a complementary discipline at the harness level.

Validating / contributing

python3 scripts/validate_plugin.py checks the manifests and skill frontmatter. See CONTRIBUTING.md.

License

MIT — see LICENSE.

Ver o original no GitHub

Bom para

Relacionados