Review And Link

When to use

The user has captured notes and wants to integrate them into the link graph. This is the "inbox triage" workflow — small, focused, repeated. Default horizon: last 7 days OR an Inbox/ folder if present.

Procedure

1. Call vault_info. Check if an Inbox/ (or 00 Inbox/, 0 Inbox/) folder exists. If yes, use that; if no, use list_notes(modified_since="7d").
2. If more than 15 candidate notes, ask the user how many to review this session (default 5).
3. For each candidate, in order of mtime descending:

• Call read_note(path=candidate).
• Call find_related(seed=candidate, top_k=5) and find_backlinks(path=candidate).
• If the note already has 3+ outgoing wikilinks, skip ("looks integrated") and move on.
• Otherwise: surface to user. Show the note's lead paragraph (truncated to ~200 chars), the top 3 related candidates, and any existing backlinks.
• Propose 1–2 wikilink inserts. The user picks: insert, edit, skip, or "move to MOC X."
• Apply the chosen action via link_notes(..., confirm=true).

1. After the session, summarize: "Reviewed {N} notes. Inserted {K} links. {M} still untouched."

Output format (per note)

Reviewing: [[Inbox/2026-05-12 Reranking trade-offs]]
Lead: "Cross-encoders dominate on top-100 recall but cost 10× the latency of bi-encoders..."

Top related (not yet linked):
  1. [[Bi-encoder vs cross-encoder]] — sim 0.78
  2. [[Latency budget for retrieval]] — sim 0.71
  3. [[MOCs/RAG research]] — sim 0.64

Proposed insert: `[[Bi-encoder vs cross-encoder]]` near "Cross-encoders dominate..."
Action? (insert / edit / skip / move-to-MOC)

Anti-patterns

• Don't batch more than 5–10 notes per session. The whole point is supervised one-at-a-time review.
• Don't propose links to notes the user has already explicitly skipped this session. Track session-state in your chat reply.
• Don't auto-archive (move to a different folder) without explicit instruction. Linking ≠ archiving.
• Don't summarize the note for the user — they wrote it, they remember it. Just show the lead paragraph and the candidates.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: bettyguo
• Source: bettyguo/obsidian_mcp
• License: MIT

Install and usage instructions live in the source repository linked above.