convez.ai
SkillDevOps & InfraestruturaAvançado
GuidanceStudio

Code Repository Audit Skill

GuidanceStudio/code-repository-audit-skill

Auditorias de código severidade-calibradas, não genéricas

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill Claude para auditorias de código metodológicas e contextualizadas. Usa framework de 13 dimensões, roteamento por intenção (PR rápida, security, pré-release, deep, full) e ferramentas OSS integradas. Stack-aware para PHP/Laravel, Python/FastAPI, TypeScript/Node, shell e Docker.

Instalação

$npx convez-ai GuidanceStudio/code-repository-audit-skill

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
0
Forks
0
Acompanhando
0
Saúde
37%
Último commit
há 2 semanas
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Revisar PRs com achados severidade-classificados
  • Security audits OWASP-flavoured em serviços críticos
  • Validação pré-release com tech-DD estruturado
  • Análise profunda de dimensões específicas (D4, D5, D7)
  • Auditorias trimestrais com relatório estruturado

Veredito

Skill produção-ready que resolve o problema real de reviews genéricos e alarmistas. Dogfooded, framework bem-fundamentado em OWASP e padrões multi-stack; instância forte para times que precisam auditar código com rigor sem ruído.

README

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

code-repository-audit-skill

A Claude skill for honest, actionable software audits. Routes by intent (quick PR review, security pass, pre-release check, deep dimension, or full tech-DD) across a 13-dimension framework.

Built for fast-but-solid startups. Generalist core, stack-aware extensions for PHP/Laravel, Python/FastAPI, TypeScript/Node, shell, and Docker.

Why this exists

Generic "review this code" prompts produce three failure modes:

  • Soft — "looks good, here are 3 style nits" on a PR with a real auth bug.
  • Alarmist — every finding is 🔴 CRITICAL, severity dilutes.
  • Generic — output reads the same regardless of stack, repo, or product context.

This skill replaces all three with methodical, severity-calibrated, context-aware reviews. It absorbs the best of:

…and adds dimensions and patterns that AI-assisted-code workflows specifically need: setup replicability, LLM-clone density, regression-after-fix discipline, async-pool tenant-context leaks.

Install

Local (clone + install)

git clone git@github.com:GuidanceStudio/code-repository-audit-skill.git
cd code-repository-audit-skill
./install.sh

Remote (one-liner)

bash <(curl -fsSL https://raw.githubusercontent.com/GuidanceStudio/code-repository-audit-skill/main/install.sh)

The installer copies claude/code-review/ to ~/.claude/skills/code-review/. Pass --force to overwrite an existing installation.

Use

In Claude Code, invoke any of:

/code-review
"review this PR"
"audit my codebase"
"security review"
"is this ready to ship?"

The skill picks the right cut based on your phrasing, or asks if it's ambiguous:

CutWhenOutputEffort
quick"review this file/PR"Inline findings table on the diff5-10 min
security"security review", "OWASP"OWASP-flavoured pass, D4 + D5 + deps30 min
release"ready to ship?"Per-release-tag gate: scan all dims, deep on diff since previous tag30 min
deep"deep audit on D5"One or more dimensions, full method run1-2 h per dim
full"full audit"Whole 13-dim report, written to docs/internal/tech-audit-YYYY-MM-DD.md5-7 h

Example sessions

Quick PR review:

You: /code-review on https://github.com/our-org/our-repo/pull/143

Claude: [scans the diff, picks dims based on what was touched, emits a
quick-scan output with severity-tagged findings + ship/hold verdict]

Security pass on a service:

You: security review on the gateway/ folder

Claude: [runs D4, D5, D7-security-subset; loads matching threat-models;
emits a security-pass output with top risk + ranked findings + tool output
summary + ship/hold recommendation]

Full audit, quarterly:

You: full audit of this repo, write the report

Claude: [walks all 13 dimensions per cuts/full.md; writes
docs/internal/tech-audit-2026-06-15.md with executive summary, status
overview, per-dim findings, and proposed follow-up milestones]

The 13 dimensions

#TitleTag
D1Code essentialityalways-deep
D2Docs integrityalways-deep
D3Tests as adversariesalways-deep
D4Security posture⚠️ default-deep
D5Multi-tenant isolation⚠️ default-deep
D6Operational readinessscan
D7Dependency hygienescan
D8Build / CI / dev-loop⚠️ default-deep
D9Data model integrityscan
D10Performance & cost baselinerelease-only deep
D11Legal / compliancescan
D12Admin surface consistencyscan
D13Setup replicabilityalways-deep

D1, D2, D3, D13 are always-deep — they decay invisibly between audits. Nobody notices code rot, doc drift, test homogeneity, or replicability regression until someone tries to clone fresh and the build fails for reasons that weren't in the README.

D4, D5, D8 carry the ⚠️ default-deep mark because they're the high-blast-radius dimensions for B2B SaaS — auth, multi-tenant, CI.

Cadence (what to run when)

TierCutsWhen
Per-push CI(automated tools only)Every PR + push to main
Per-release-tagrelease cutBefore tagging v0.x.y (~30 min human)
Quarterlyfull cut + restore-from-backup drillEvery 3 months (~5-7 h)
Triggereddeep, security cutsNew integration; post-incident; pre-VC

See claude/code-review/playbooks/operations.md for the full cadence + 5-week starter pack + anti-patterns to avoid + things-that-bite-if-skipped.

Severity scheme

MarkMeaningExample
🔴Ship-blocker — fix before next audit passSecret in git history; cross-tenant data leak; SQLi in admin panel
🟡Fix soon — defensible to leave; real auditor will noteHealthcheck missing; dep 6mo behind; doc claim mismatches code
🟢Nice-to-have — acceptable, document, move onTODO in dev tooling; non-pinned base image tag

Per-project extensions

Drop .code-review/extras/*.md files in your repo to add project-specific dimensions, threat models, or stack files. The skill loads them after the defaults. See claude/code-review/extensions/README.md for the schema.

Repository layout

code-repository-audit-skill/
├── install.sh                       # installer
├── README.md                        # this file
├── LICENSE                          # Apache-2.0
└── claude/
    └── code-review/                 # the skill that gets copied to ~/.claude/skills/
        ├── SKILL.md                 # routing spine — Claude loads this first
        ├── routing/detect-stack.md
        ├── cuts/{quick,security,release,deep,full}.md
        ├── dimensions/D{01..13}-*.md
        ├── languages/{php-laravel,python-fastapi,typescript-node,shell,docker}.md
        ├── threat-models/{multi-tenant-isolation,secret-management,
        │                  llm-assisted-code,pii-data-flow,auth-model,
        │                  data-loss}.md
        ├── tools/{_matrix,gitleaks,trivy,semgrep,phpstan-larastan,
        │         ruff,mypy,biome,hadolint,jscpd,pgtap,auto-dep-updates}.md
        ├── templates/{full-audit-report,quick-scan-output,
        │             security-pass-output,triage-and-summary,
        │             finding-phrasing,ci-github-actions.yml,
        │             ci-pre-commit.yaml}
        ├── playbooks/{operations,false-positives}.md
        ├── extensions/README.md
        ├── examples/{full-audit-multitenant-saas,security-pass-fastapi}.md
        └── scripts/{_detect_stack,_bootstrap_tooling,
                    _doctor_fresh_clone,_findings_to_milestones}.py

Each file has one clear purpose. Files cross-reference rather than duplicate — updating a fact happens in one place. The skill itself follows the rules it audits (D1 + D2).

Design principles

  1. Lazy load — the spine (SKILL.md) is ~250 lines; sub-files are loaded only when their cut / dimension / language matches. A quick cut on a TypeScript PR loads ~5-7 files, not all 40+.
  2. Generalist core, stack-aware extensions — dimensions are stack-agnostic; languages/ carry stack-specific gotchas.
  3. Provenance tracked — each concept notes its source (Anthropic skill, OWASP, internal framework). Future updates traceable.
  4. No bloat — every file earns its place. Files >300 lines are split. The skill eats its own dogfood on D1.
  5. Severity discipline — 🔴 / 🟡 / 🟢 calibrated to "what would a real auditor say at a VC pitch?". No alarmism, no padding.
  6. Output template-driven — every cut has a matching template. Reports are consistent across audits + comparable over time.

Anti-patterns this skill explicitly rejects

  • SOC2 readiness program at PoC stage. Wait for first enterprise prospect.
  • PR review with 2+ approvers for 1-3 engineers. Strong CI is the reviewer.
  • 100% line-coverage targets. Branch coverage on tenancy + billing only.
  • External pen-test before product-market fit. €5-15k for findings CI would catch.
  • Snyk / SonarCloud paid tier. Trivy + Semgrep + native deps audits cover ~90% for €0.
  • STRIDE threat-model documents per feature. Nobody reads them. A 5-line release-time "what changed in trust boundaries?" note works.

See claude/code-review/playbooks/operations.md § Anti-patterns for the full list.

Contributing

Open an issue or PR against GuidanceStudio/code-repository-audit-skill.

A finding that recurs across projects might warrant a default dimension or threat-model — propose it. A stack-specific gotcha that isn't in languages/ — add it. Keep additions in the same style: short, concrete, cross-referenced.

License

Apache-2.0. See LICENSE.

Built on the prior art of the projects listed in Why this exists; each retains its own license. The skill's own framework + glue + dimensions are Apache-2.0.

Ver o original no GitHub

Bom para

Relacionados