convez.ai
SkillAutomaçãoIntermediário
DatJavaClass

Diamond QC

DatJavaClass/diamond-qc

Controle de qualidade multi-agente adversarial antes de publicar

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill de controle de qualidade adversarial multi-agente para Claude Code que implementa um protocolo de verificação em camadas: autor valida claims por execução, dois auditores decorrelacionados caçam erros por ângulos diferentes (fidelidade e execução), e um meta-auditor captura pontos cegos compartilhados. Verifica por evidência, não por assertion.

Instalação

$npx convez-ai DatJavaClass/diamond-qc

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
42%
Último commit
há 4 meses
Versão
v1.0.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Revisar código antes de deploy em produção
  • Validar análises de dados e números críticos
  • Verificar correção de documentação técnica e research
  • Auditar configurações e infraestrutura
  • Revisar conteúdo escrito antes de publicação

Veredito

Diamond QC é uma abordagem sofisticada e bem fundamentada para QA através de decorrelação de agentes, não repetição. Funciona particularmente bem para artefatos com claims verificáveis (código, dados, análises) onde evidência executável é disponível.

README

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

Diamond QC

A small Claude Code skill for catching errors in an artifact before it ships: study notes, code, prose, research, configs, data; anything with claims you can actually check.

It's named for the shape: one author at the top, two independent auditors across the middle, one meta-auditor at the bottom point. The bottom point is the "who watches the watchmen" step.

Why it's adversarial (and agentic)

A single reviewer (including the author) shares blind spots with itself. Re-reading your own work mostly re-confirms it. Diamond QC fixes that by spreading the check across several independent agents that are deliberately aimed at the artifact from different angles:

  • Author drafts, then verifies every claim it can by execution: runs the code, computes the numbers, traces each fact to a source. Known-shaky work is fixed here, not passed downstream.
  • Auditor A, fidelity. Does every claim match the source/spec/reality? Hunts fabrications, distortions, and unsupported assertions, quoting exact text.
  • Auditor B, execution. Runs the code, runs the numbers, checks types, shapes, boundaries, edge cases. "I checked it" is not allowed; it shows the run.
  • Meta-auditor, watches the watchmen. Catches what both auditors missed (their shared blind spot) and overturns what an auditor got wrong (false positives), then delivers the final verdict and the minimal set of fixes.

The point is decorrelation, not repetition. Two auditors running the same check share the same blind spot and waste a seat. Diversity of attack is what catches errors. The skill verifies by evidence, not assertion (recall is not verification) and escalates effort by risk, not by running the same pass twice.

It's also right-sized: a tiering rule (SKIP / LIGHT / FULL) keeps you from spinning up five subagents to check a one-line fix.

What you need

  • Claude Code installed.
  • That's it. The skill is a single Markdown file. The auditors run as Claude Code subagents, so no extra dependencies, API keys, or services.

Install

Clone (or copy) the skill into your Claude Code skills directory.

User-level (available in every project):

git clone https://github.com/DatJavaClass/diamond-qc.git ~/.claude/skills/diamond-qc

Project-level (this repo only):

git clone https://github.com/DatJavaClass/diamond-qc.git .claude/skills/diamond-qc

On Windows the user-level path is %USERPROFILE%\.claude\skills\diamond-qc.

Claude Code discovers a skill by the SKILL.md at the root of its folder, so the folder must be named diamond-qc and contain SKILL.md directly.

Use

Just ask Claude Code to run it, or name it directly:

  • "Diamond QC this before I send it."
  • "Run the Diamond on this analysis."
  • "Who watches the watchmen, verify these numbers."
  • Or any request to rigorously verify / fact-check / QC something before it ships.

Claude picks the tier (SKIP / LIGHT / FULL) based on stakes and tells you which it chose. See SKILL.md for the full protocol.

License

MIT © Written by DatJavaClass (Victor S.) Edited by Claude, driven insane by ChatGPT, made to cry by Grok

Ver o original no GitHub

Bom para

Relacionados