AI智能体协作知识库模板

Obsidian + Claude Code starter

An Obsidian vault template built for working with coding agents. You keep your projects, people, meetings and research as plain Markdown; Claude Code (or Codex, Cursor, and other agents that read AGENTS.md) works inside the same folder, following rules and workflows that live in the vault itself.

Everything is plain files in git. No database, no server, no plugins required.

Who it is for

• You use (or want to use) Obsidian as a personal knowledge base, and you want an AI agent that can file, find and update notes without you explaining your system every session.
• You are comfortable opening a terminal and running one command. You do not need to be a programmer.

What you get

• A layout with rules agents can follow: Inbox/, Projects/, People/, Meetings/, Knowledge/, Journal/, Archive/, System/, each note typed by frontmatter (vault schema).

• One instruction file for every agent: AGENTS.md holds the vault's rules; CLAUDE.md imports it, so Claude Code, Codex and others share one copy.

• Eight skills (slash commands in Claude Code):

  Skill	What it does
  /onboarding	Interviews you, fills in "About me" in AGENTS.md, creates your first notes
  /inbox-triage	Files everything in Inbox/ into the right project, person or research note
  /meeting-notes	Turns rough notes or a transcript into a meeting note and updates the people and project notes; also preps you for a meeting
  /project-brief	Starts a project with a real definition of done, or catches you up on an existing one
  /decision-record	Records a decision with its options, reasons, and when to revisit it
  /research-capture	Turns a link, paper or conversation into a sourced research note
  /weekly-review	Drafts a weekly review from what actually changed and what is still open
  /log-feedback	When you correct the agent, fixes the output and the skill or rule that caused it

• Templates for every note type in System/Templates/, wired to Obsidian's core Templates and Daily notes plugins.

• A session-start hook that tells the agent what is waiting (inbox items, active projects, uncommitted changes), and a statusline.

• A vault checker, scripts/check_vault.py, that finds broken links, missing headings and untyped notes.

• A feedback loop: System/agent-feedback.md collects friction, and the weekly review folds fixes back into the skills and rules.

• Fictional example notes in _examples/ folders showing each pattern, removable with one command.

Quickstart (5 minutes)

You need git, Obsidian, and Claude Code (curl -fsSL https://claude.ai/install.sh | bash on macOS or Linux; it needs a Claude subscription or API account).

git clone https://github.com/Rome-1/obsidian-claude-starter.git my-vault
cd my-vault
./setup.sh --fresh-git      # checks tools, starts your own git history

--fresh-git drops this template's history and remote so your notes start in a clean repository of their own. Leave it off if you would rather keep the history.

Then:

1. Open it in Obsidian: "Open folder as vault" and pick my-vault.
2. Start Claude Code in the same folder: claude. The first time, Claude Code asks whether to trust the folder. Say yes; that turns on the vault's settings, hook and skills.
3. Type /onboarding. It asks a few questions, personalizes AGENTS.md, and creates your first project and person notes.
4. When you no longer need the examples: ./setup.sh --remove-examples.

To back the vault up, create a private repository on GitHub (or anywhere) and push to it:

git remote add origin <your-private-repo-url>
git push -u origin HEAD

Day to day

• Drop anything into Inbox/ (from Obsidian, your phone, or by asking the agent). Run /inbox-triage when it piles up.
• After a meeting, paste your notes and run /meeting-notes. Before one, ask "prep me for my call with ...".
• Before working on a project, ask the agent to catch you up (/project-brief); it reads the brief and current state first.
• Friday: /weekly-review.
• When the agent gets something wrong, say so and run /log-feedback, so the fix lands in the skill and not just in the note.

Customize

Rename or add folders freely; update the layout table in AGENTS.md and the schema when you do, so agents know.

Optional add-ons

• Vault search from other projects: an MCP server so agents working in your code repos can look things up in your notes. ./setup.sh --mcp, or see System/setup-mcp.md.
• beads task tracking: a dependency-aware issue tracker for agents, if checkboxes stop being enough. See System/beads.md.
• Obsidian plugins: Dataview, Obsidian Git, Templater. See System/recommended-plugins.md.

Layout

.
├── AGENTS.md              # agent instructions for this vault (the one to edit)
├── CLAUDE.md              # imports AGENTS.md for Claude Code
├── setup.sh               # setup and checks (macOS, Linux)
├── Inbox/                 # raw captures
├── Projects/              # one file per project, or a folder: Brief, Current-State, Decisions
├── People/                # one note per person
├── Meetings/              # YYYY-MM-DD-Topic.md
├── Knowledge/             # research notes and references
├── Journal/               # daily notes, weekly reviews
├── Archive/               # finished or dropped material
├── System/
│   ├── index.md           # map of the system notes and skills
│   ├── vault-schema.md    # note types, frontmatter, links
│   ├── writing-guide.md   # how prose should read
│   ├── agent-feedback.md  # friction log
│   ├── global-agent-rules.md
│   ├── claude-code.md     # permissions, hooks, statusline, skills
│   ├── setup-mcp.md       # optional
│   ├── beads.md           # optional
│   └── Templates/
├── scripts/check_vault.py
└── .claude/
    ├── settings.json
    ├── hooks/session-context.sh
    ├── statusline.sh
    └── skills/            # the eight skills above

Design choices

• Markdown in git, not a database. Notes stay readable in any editor, diffs show what an agent changed, and history is one git log away.
• Rules live next to the notes. An instruction you will need again goes in AGENTS.md, a skill, or a template, not in a chat that disappears.
• Short always-loaded context. AGENTS.md is loaded every session, so it stays short; procedures load only when their skill runs.
• Verify the result, not the exit code. The checker exists so "I created the note" can be confirmed, and the skills tell agents to run it.
• Ask before anything irreversible. Deleting, bulk moves, and pushing need your yes.

Contributing

Issues and pull requests are welcome, especially new skills that are general enough to help most people. Please keep examples fictional and generic.

License

MIT