// the find
eugeniughelbur/obsidian-second-brain
Cross-CLI skill for Obsidian: turn your vault into a living AI-first second brain across Claude Code, Codex, Gemini, OpenCode, Hermes, and Pi. 44 commands - self-rewriting notes, local+hybrid semantic search, key-less web research, /obsidian-architect codebase docs, and scheduled agents that maintain the vault while you sleep.
A Claude Code skill (and multi-platform adapter) that turns an Obsidian vault into an AI-managed knowledge base. Instead of appending notes, it attempts to rewrite existing pages when new information arrives — contradictions get reconciled, people pages get updated, patterns get synthesized. Aimed at knowledge workers who spend serious time in Obsidian and already live in a Claude Code terminal.
The vault-first research flow in /research-deep is the genuinely smart idea here: it scans what you already know before spending API tokens on external queries, and only fills gaps. That's not obvious and it compounds over time. The semantic search implementation is done right — optional, fuses with keyword fallback, runs locally via Ollama so notes never leave the machine, and degrades silently rather than breaking. The multi-platform build system (build.sh emitting platform-specific adapters for Codex, Gemini, OpenCode, Hermes) is real engineering, not just duplicated markdown. Bi-temporal fact tracking — recording when something was believed true versus when the vault learned it — is the kind of detail most personal knowledge tools skip entirely.
The core promise — vault pages 'rewrite themselves' — is entirely LLM instruction-following with no transaction safety. If /obsidian-ingest decides to reconcile a contradiction incorrectly, there's no rollback, no diff, no audit trail beyond a log.md entry. The scheduled agents running claude -p headlessly at 10 PM to modify your vault while you sleep is a data integrity risk that the README buries. The 44-command count includes a lot of thin API wrappers (Perplexity, xAI/Grok, Gemini File Search, YouTube Data API) that require separate paid keys — the 'free, key-less fallback' for /research is Wikipedia and HackerNews scraping, which is fine but not what the feature table implies. The multi-platform story also overpromises: the Codex and Gemini builds are static instruction files; whether the synthesis-heavy commands (/obsidian-architect, /obsidian-reconcile) actually work reliably on non-Claude models is an open question the README hedges on rather than answers.