obsidian-kb

Obsidian Personal Knowledge Base

Installed compatibility resources live at ~/.obsidian-kb-skill/skill; use that directory as <skill-root> for references, assets, and bundled helpers.

Overview

Agent-agnostic instructions to create, organize, and update notes in an Obsidian vault. Any tool that reads/writes local files can use it.

DO NOT auto-save

This skill never writes to the vault on its own. Write only after explicit save intent ("save to Obsidian", "沉淀", "总结存档"). An explicit request to evaluate what from a conversation is worth capturing may run conversation-harvest.md as analysis only. For unrelated Q&A, debugging, or casual chat: do nothing.

When the user asks to save or review capture candidates

1. Route analysis before Vault discovery: for an analysis-only conversation review, read conversation-harvest.md, return the candidate proposal, and stop without locating or scanning a Vault. Otherwise continue.
2. Find vault for a write: env OBSIDIAN_KB_VAULT → ~/.obsidian-kb-config → ask. Refuse unless a real vault (.obsidian/ + Templates/).
3. Read the smallest operation-specific reference set before writing. Read Vault governance first, then let one compact discovery call answer this: required_references names every file the selected type, template, and destination require, so read that set instead of discovering it one failure at a time. Pass the governed route as --folder — a crowded child folder does not make its parent look crowded. The full map:
   • New note: read only note-creation.md.
   • Finished source-backed article: after note-creation.md routes it, also read web-capture.md; verified captures additionally load deep-capture.md.
   • Crowded selected destination: when compact discovery reports it, also read folder-routing.md; uncrowded destinations do not load it.
   • Existing note: read only update-note.md.
   • Filing the Inbox: read only process-inbox.md. Filing moves notes the user already wrote; it never authors one, so ≤1 note written does not bound it.
   • Conversation context archive: read conversation-digest.md.
   • Conversation knowledge review: read conversation-harvest.md; load note-creation.md only after one durable candidate is selected for writing.
   • Task Memory: read task-memory.md only after explicit opt-in (off by default).
   • YAML, rules, and Git references are troubleshooting or post-processing material: load yaml-standards.md, rules-and-errors.md, or git.md only when the current task requires it.
4. Prefer bundled helpers (never a one-off script): run python <skill-root>/scripts/run_helper.py <helper> ..., where <skill-root> contains this SKILL.md. For new notes use create-note --preflight-json, inspect the structured validation, then apply with --from-preflight <content.sha256> --apply --compact-json instead of resending the body; update-note is only for Task Memory, while ordinary existing-note edits follow update-note.md with native file tools.
5. Stay bounded: ≤10 files scanned, ≤1 note written, ≤5 wikilinks. Never overwrite; add -2 on name clash. Validate after.
6. Route: daily→15-Daily · meeting→10-Work · learning/article→20-Learning · insight/digest→30-Insights · project→40-Projects · person→50-People · quick/unread source→00-Inbox. An ordinary finished article uses standard capture; verified capture is explicit or evidence-sensitive.