Topics You Didn’t Tag
Unlike folders and tags, topical clusters aren’t labels you assigned — they emerge from how your ideas actually co-occur. That’s why the first look at your own notes is often surprising: the graph shows the themes as they are written, not as they’re filed.
A topic you consider central may be peripheral; two topics you keep separate may already overlap.
Open your Step 1.1 graph and look at the Analytics panel:
- Main topical clusters — each color is a group of concepts that appear in the same context, with a relative influence percentage. These are your themes.
- Main concepts — ranked by betweenness centrality: the ideas the rest of your discourse routes through. These aren’t necessarily the most frequent words — they’re the most connective ones.
- Topical summary — the built-in AI can write a short outline of each cluster, turning the graph into a structured summary of your text you can reuse as a content outline.
How to Read Them
- The two or three biggest clusters are your dominant narrative. If one of them towers over the rest, note it — Step 3 will tell you whether that’s a bias.
- The smallest clusters often hold the most original thinking — ideas you touched once and never developed. Don’t prune them; they’re raw material for Steps 2.2 and 2.3.
- Select any cluster to see only its statements — an instant thematic reading of your own text (“an upgraded tag cloud” — see the full tutorial).
- Discourse entrance points — concepts with high influence per occurrence — are the easiest doors into the discourse: good candidates for titles, sections, and hooks.
Same Insight, in Your LLM
With the MCP server connected, ask:
- “Analyze my graph called ‘[name]’ and list the main topical clusters with a summary of each.” (
generate_topical_clusters) - “What are the most influential concepts in this graph and what connects them?” (
generate_knowledge_graph)
The LLM receives the same structure you see on screen — clusters, concepts, relations — so its summary is grounded in your actual discourse rather than generic training data (why that matters).