obsidian-visualize

Use the obsidian-second-brain skill. Execute /obsidian-visualize $ARGUMENTS:

The optional argument is a scope: an EXACT note title or alias (a project, an entity), or "full" for the entire vault. Default: full vault. A free-form topic phrase will not match - resolve it to a real note title first (e.g. via /obsidian-find).

1. Read _CLAUDE.md first if it exists in the vault root

2. Build the graph deterministically with the scanner instead of reading every note into context (a full-vault read is O(read-everything) and burns the budget). Run from the skill root (its absolute path was given at session start as Skill root; substitute it for SKILL_ROOT):

   uv run --directory "SKILL_ROOT" scripts/link_graph.py --path "<vault>" [--scope "<topic/project/entity>"]
   

   It returns JSON with nodes (path, title, type, folder, in/out/degree), edges (resolved [[wikilink]] pairs, each {from, to, type: "link"}), typed_edges (the semantic overlay from relations: frontmatter - {from, to, type} where type is supersedes, depends_on, caused, etc. per references/ai-first-rules.md Rule 6), typed_edge_problems (unhonored typed edges), and stats (node_count, edge_count, typed_edge_count, orphan_count, dangling_link_count, top_hubs, orphans). Pass --scope with an exact note title/alias (the script keeps that note plus its 2-hop neighborhood); omit it for the full vault. If the result has node_count: 0, the scope did not resolve to a note - find the real title and rerun. Use this JSON as the graph - only open individual notes if you need a label the scan did not provide.

3. Generate a JSON Canvas file (.canvas) compatible with Obsidian's native canvas viewer:

   Structure:

   {
     "nodes": [
       {"id": "1", "type": "file", "file": "wiki/entities/Ada Lovelace.md", "x": 0, "y": 0, "width": 250, "height": 60},
       {"id": "2", "type": "file", "file": "wiki/projects/Tide Gateway.md", "x": 300, "y": 0, "width": 250, "height": 60}
     ],
     "edges": [
       {"id": "e1", "fromNode": "1", "toNode": "2"}
     ]
   }
   

   Layout rules:

   • Hub nodes (most links) go in the center, larger
   • Cluster by type: entities on the left, projects top-right, concepts bottom-right, daily notes bottom
   • Color by type: entities = blue, projects = green, concepts = purple, daily = gray, sources = orange
   • Edge labels = when two nodes connect through multiple links, record the count in the edge label (JSON Canvas edges have no thickness property). When a typed_edges entry connects the same pair, label the edge with the relation type instead (e.g. supersedes, depends_on) - a typed edge is more informative than a count
   • Orphan nodes placed at the edges with a red border (easy to spot)

4. Save to vault root as atlas.canvas (or atlas-{topic}.canvas if scoped)

5. Also generate a text summary with centrality ranking (use the scanner's stats and per-node degree/in/out directly - do not recompute by hand):

   • Total nodes and edges (stats.node_count, stats.edge_count), plus the dangling_link_count (wanted notes - links to unwritten notes).
   • Typed edges - if stats.typed_edge_count > 0, summarize the semantic overlay: count by relation type (from typed_edges), name the longest supersedes/depends_on/caused chains (these are the reasoning paths flat links can't express), and flag any typed_edge_problems as a nudge to run /obsidian-health for the full lint. If it's 0, note that the vault has no typed edges yet and point to references/ai-first-rules.md Rule 6 as the way to add them.
   • Hub nodes (centrality) - top 5 from stats.top_hubs (already ranked by degree), each with its link count and a one-line "everything flows through this because..." note. A hub qualifies if its degree is at least 3x the median, or it sits in the top 1% of the vault - whichever surfaces fewer.
   • Bridge nodes - nodes that, if removed, would split a cluster. Rank by betweenness (approximate: count the shortest paths each node sits on between the top-10 hubs). These are the load-bearing connectors; surface the top 3 with the two clusters each one joins.
   • Orphan nodes - no connections, listed by type. Flag any that are >30 days old (stale orphans are higher-priority cleanup targets than fresh ones).
   • Clusters - groups of tightly connected notes, named by their hub. Note any cluster with <3 cross-cluster edges (those are silos).
   • Centrality skew - if one node holds >25% of total edges, call it out as a single point of failure for navigation.