convez.ai
SkillDevOps & InfraestruturaIntermediário
vindm

dotclaude — Framework de Infraestrutura IA para Claude Code

vindm/dotclaude

Metodologia estruturada de dev infrastructure via entrevista hierárquica automatizada

Ver no GitHub
InicialPluginScore · de 0 a 100

Plugin que bootstrapa a infraestrutura de desenvolvimento do seu projeto através de uma entrevista hierárquica em 7 camadas, gerando automaticamente CLAUDE.md, documentação e configurações .claude/ otimizadas. Implementa metodologia de produção testada, com kits especializados para design, coding, testing, dados e workflows com LLM.

Instalação

$npx convez-ai vindm/dotclaude

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
57%
Último commit
há 2 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Onboarding de novos projetos com Claude Code com arquitetura clara desde o início
  • Documentação versionada e knowledge graph autogerado para context window eficiente
  • Padronização de processos, quality bars e discipline hooks em equipes

Veredito

Framework sólido que sistematiza o setup de projetos para Claude Code, evitando decisões ad-hoc. A abordagem em 7 camadas é rigorosa mas exige investimento inicial; ideal para projetos que priorizam contexto bem estruturado e escalabilidade.

README

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

AI dev infrastructure framework for Claude Code.

Walks you through 7 layers — project identity → architecture → process → quality bar → knowledge graph → domain kits → maintenance — and authors CLAUDE.md + docs/ + .claude/ derived from your project.

Not templates. Not a kit. A methodology.

Claude Code License Version

claude plugin marketplace add vindm/dotclaude
claude plugin install dotclaude@dotclaude
# Then in any project root:
/dotclaude:bootstrap

dotclaude — AI dev infrastructure for Claude Code

A Claude Code plugin that bootstraps your project's CLAUDE.md, docs/ knowledge graph, and .claude/ system through a hierarchical interview. Distilled methodology from one battle-tested production codebase.

Jump to: Slash commands · How it works · See it in action · What to expect · FAQ · Install

Slash commands

CommandWhat it sets up
🪄 /dotclaude:bootstrapHeadline. Full 7-layer hierarchical interview. Authors CLAUDE.md + docs/ + .claude/.
🎨 /dotclaude:designShowpiece. UI / IA / a11y / visual-quality audit kit (Layer 6 standalone).
🧹 /dotclaude:codingFile-size discipline, code review, voice / forbidden phrases.
📐 /dotclaude:planningPre-impl validation, audit routing.
🧪 /dotclaude:testingTest architecture + coverage strategy.
🗃️ /dotclaude:dataDB integrity, query discipline, migrations.
🤖 /dotclaude:ai-workflowLLM cost monitoring + eval discipline.
🪄 /dotclaude:initv1 entry — Layer 6 only, skips upstream layers.

Inside the plugin — 35 methodology principles · 12 hook templates · 7 domain skills + 1 bootstrap orchestrator · 4 anonymized war stories. Cost: ~984 tokens always-on per session; 2.7–5.5k per skill invocation; ~12k for full bootstrap. Full architectural reference in docs/v2-vision.md.

How it works

The 7-layer hierarchy

Each layer authors a specific slice of CLAUDE.md + docs/ + .claude/. Dependency-ordered — Layer 1's output feeds Layers 2–6; Layer 7 polices the rest.

#LayerWhat bootstrap authors
1Project IdentityCLAUDE.md opening — vision + ICP + moat + production-vs-internal + stage
2ArchitectureCLAUDE.md Architecture + Constraints + boundary rules + universal hooks
3Process DisciplineCLAUDE.md "How You Work" + DoD + task-classification table + verification ladder + memory typing
4Quality Bar.claude/rules/design-north-star.md (Tier 1 + Tier 2 benchmarks + anti-references) + quality-bar skill
5Knowledge Graphdocs/README.md + subdirectory skeleton + docs/product/capabilities.md scaffold
6Domain KitsPer applicable domain: design / coding / planning / testing / data / ai-workflow. .claude/{agents,skills,hooks,rules}/*
7MaintenanceSaturday-style design-debt ritual + drift detection + skill-auditor agent (default-deferred for early projects)

Bootstrap in 5 phases

PhaseWhat happensDuration
1. Project scanAuto-discovers ~30–40% of inputs from package.json, file tree, git log, existing CLAUDE.md / docs/ / .claude/~5s
2. Hierarchical interview7 phases (A–G), 1–3 Qs per turn, skip-if-Phase-1-answered20–60 min
3. Cross-layer coordinationMerges forbidden-phrase lists; reconciles audit routing< 1 min
4. Stage → review → commitAll authoring lands in .claude-staging/ first; explicit approval gatevaries
5. Output summaryInventory by layer + skipped reasons + next-step recommendations< 30s

Brownfield safety

Bootstrap detects your project's state via Phase 1 scan and runs in the right mode automatically.

Project stateModeWhat bootstrap does
No CLAUDE.md + no .claude/FreshRuns all 7 layers
CLAUDE.md < 50 LOC OR .claude/ < 5 artifactsAPPENDAdds missing layers only; never stomps existing
CLAUDE.md > 200 LOC OR > 5 H2s + structuredREFUSERecommends Layer 6 standalone or manual edits

REFUSE is a feature — it protects accumulated AI infrastructure from being overwritten.

See it in action

/dotclaude:design running on a fresh Vite + React + TypeScript project, authoring 8 tailored design artifacts (4 agents + 2 skills + 2 hooks) in ~15 seconds:

Validation reports (real runs, not simulations):

What to expect

Wall-clock per project shape:

Project shapeBootstrap time
Greenfield, 1–3 domains20–35 min
Early prototype, 3–5 domains40–60 min
Shipped, 5–7 domains60–90 min
Mature, 8 domains (max-domain)90–120 min
Brownfield comprehensive (REFUSE)< 5 min, no authoring

What bootstrap can't manufacture in one pass:

  • Accumulated incident memories — emerge from coding sessions over weeks/months
  • Per-aspect design system docs — land when content emerges from real design work
  • Substrate runbook skills — written when you've lost > 1 hour relearning a subsystem twice

Bootstrap creates the scaffold + the "accrue-here" anchors. The entries themselves accumulate over time. The methodology compounds; one-pass output is the seed, not the tree.

Full depth-match analysis: docs/bootstrap-smoke-test-2026-05-21.md.

FAQ

How is this different from a CLAUDE.md template?

Templates are static text you paste + edit. dotclaude reads your actual project (package.json, file tree, git log --grep="fix:", existing conventions) and authors a CLAUDE.md tuned to it. The 7 layers are a methodology, not a template.

Will dotclaude overwrite my existing CLAUDE.md or .claude/?

No. Phase 1 scan detects existing infrastructure. Mature setups trigger REFUSE mode (recommends Layer 6 standalone instead). Partial setups trigger APPEND mode (adds missing layers only, never stomps). All authoring goes to .claude-staging/ first; explicit approval gate before commit.

Does it work for non-iOS / non-React-Native projects?

Yes. The methodology is platform-agnostic. Smoke tested on a Vite + React + TypeScript + Tailwind project — grade A. The 7 layers map to any project shape; the domain skills (/dotclaude:coding, /dotclaude:data, etc.) are stack-universal.

Can I use just one domain skill without bootstrap?

Yes. /dotclaude:design, /dotclaude:coding, or any other domain skill runs standalone. Useful for incremental setup or adding a new concern to an existing CLAUDE.md.

What does it cost in tokens?

~984 tokens always-on per session (skill descriptions). 2.7–5.5k per single skill invocation. ~12k for the full bootstrap (loads upstream principles selectively per layer). Session overhead is < 1% of typical Claude usage.

How do I uninstall?

claude plugin uninstall dotclaude@dotclaude. The .claude/ directory dotclaude authored stays in your project — it's yours to edit, version, or remove.

Install & contribute

# Production install (any Claude Code session)
claude plugin marketplace add vindm/dotclaude
claude plugin install dotclaude@dotclaude

# Dev install (for contributing)
git clone https://github.com/vindm/dotclaude.git && cd dotclaude
claude --plugin-dir .

Contributing: full guide in CONTRIBUTING.md — covers adding principles, hooks, skills, war stories, smoke test discipline, and the PR checklist.

Changelog: see CHANGELOG.md for v1.0.0 → v1.1.0 (v2 reframe).


Built from months of working with Claude Code as a daily driver. MIT licensed · v2 architecture · changelog · contribute

Ver o original no GitHub

Bom para

Relacionados