Obsidian AI Vault Kit
MrQ-Coding/obsidian-ai-vault-kit
Cofre Obsidian gerenciado por IA com restrições, governança e deploy automático
Ver no GitHubFramework completo para manter um cofre Obsidian colaborativamente com IA através de restrições de escrita, camada de execução (skills) e governança de percepção. Integra-se com Claude Code e Codex com deploy em um comando, incluindo verificação automática de backlinks, detecção de conteúdo duplicado e sincronização inteligente de índices.
Instalação
npx convez-ai MrQ-Coding/obsidian-ai-vault-kit1 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
- Manter bases de conhecimento pessoal com assistência de IA segura
- Automatizar curadoria, indexação e detecção de problemas em vaults grandes
- Gerar resumos diários a partir de logs de conversas com IA
- Validar integridade de links e estrutura de notas continuamente
Veredito

Solução sofisticada para quem quer delegar manutenção de conhecimento a IA sem perder controle ou segurança. Requer familiaridade com Python/Node.js e arquitetura de constraints, mas oferece camada de proteção robusta (allowlists, lint, verificação de backlinks, detecção de rot). Extraído de implementação real em produção.
README
A tradução para o português deste README ainda está na fila.
Obsidian AI Vault Kit
A complete setup for handing your Obsidian vault over to AI to maintain collaboratively: write constraints + an execution-layer skill set + a perception-layer governance + one-command deploy. Works with both Claude Code and Codex.
This repo is the framework, extracted from a real, running knowledge vault — it contains no note data. The repo root mirrors the vault root: AGENTS.md / CLAUDE.md / deploy/. Copy these into a vault and you're ready to go.
What problem it solves
Letting AI read and write a personal knowledge vault directly raises four worries; this kit addresses each with a mechanism:
- Bad writes: a write-domain allowlist (each skill can only write to the directories it has declared), a do-not-read list, required frontmatter fields, and a mandatory post-write lint check.
- Messy writes: a single source of truth for index-maintenance rules; check backlinks before deleting / moving / renaming so no link is left dangling.
- Rot: a background perception-layer checkup (a dozen-plus detectors for broken links, suspected duplicates, stalled drafts, index drift, sync-conflict copies, and more); the scripts only propose candidates, the human decides.
- Saying one thing, doing another: design docs declare where they land via a frontmatter
contractsfield, andverify.pyreconciles automatically; every mechanism writes its retirement condition when it is first proposed.
Quick start
- Copy
AGENTS.md,CLAUDE.md, anddeploy/into your vault root. - Run
node deploy/deploy.mjs— installs the skills, commands, and global constraints (idempotent; just re-run after the source files change). - Run
python3 deploy/verify.py— verifies the deployment (130+ checks currently).
Prerequisites: Node.js 18+ (no npm dependencies), Python 3 + PyYAML. Full deployment notes in deploy/README.md.
What's inside
| Component | Role |
|---|---|
deploy/shared/ | Single source of truth for shared constraints: constraints.json + post-write check lint.py + backlink check backlinks.py + index rules |
deploy/obsidian-notes/ | Notes skill: find, write, organize, record routing |
deploy/obsidian-excalidraw-designer/ | Diagram skill: semantic blueprint → deterministic generator → quality gate → visual readback |
deploy/daily-digest/ | Daily digest: extracts the day's work from AI conversation logs and generates a first-person digest (mergeable across machines) |
deploy/vault-curator/ | Perception layer: content lifecycle + health checkup + memory consolidation + self-evolution, four jobs behind one entry point |
deploy/deploy.mjs | One-command deploy (8 steps), injecting the constraints into both platforms' global config |
deploy/wiki-git.sh | Optional: a local git undo layer (the git dir lives outside the cloud-sync scope) |
Design principles
- Scripts surface facts, AI judges meaning: deterministic work like scanning, validating, and rendering belongs to scripts; semantic judgment belongs to the model; the boundaries are hard-coded into the flow.
- Single source of truth: constraints live only in
deploy/shared/constraints.json, generated and injected at deploy time; no second mirror is allowed. - Opportunistic triggers, no timers: due reminders for the checkup and for catching up on the daily digest ride along at the tail end of a skill, zero-config across platforms.
- What you claim is what you check: with
contractsdeclarations and the mechanism retirement-condition table, every new mechanism writes down how to verify it and when it should be cut, right when it's proposed. - AI only proposes, the human decides: governance actions, rule changes, and git commits are never executed automatically.
Adapting it to your own vault
deploy/shared/constraints.json: do-not-read files, each skill's write domain, checkup-ignored directories — edit these to match your directory structure.- The directories referenced in
AGENTS.md(such as the source vault'svault-design/design-notes folder) follow the original vault's directory conventions; swap in your own paths. - Conventions like the daily-digest landing spot (
00-daily/YYYY/MM/MM-DD/) and the checkup report location are documented in each component's SKILL.md / README.