Turn Your Vault Into a Knowledge Graph Claude Can Traverse
TL;DRLink your notes consistently, then ask Claude to read them and write a simple map — an index of what connects to what. From there you stop searching and start asking relationship questions: your AI walks the links instead of reading files one by one.
…
Do I need special graph software for this?
No. The 'graph' is just a text file — an index or a small JSON — that Claude writes and reads. No database, no plugin, no separate app. Plain files, same as the vault itself.
What if my old notes aren't linked to each other yet?
Point Claude at the whole vault and ask it to read every note and add [[wikilinks]] and tags where notes are actually related. That's the raw material the graph is built from — do this pass first.
Isn't this just Obsidian's built-in graph view?
Different thing. The graph view is a picture for you to look at. This is a map Claude can read back and reason over — it can actually answer 'how does X connect to Y', not just draw a pretty web of dots.
How often do I need to rebuild the map?
Don't rebuild — update. Ask Claude to refresh the map any time you add or link notes, same habit as keeping index.md current in episode 01. A minute of upkeep, and it never goes stale.
Episode 01 got you a vault: notes as plain Markdown files, Claude reading and writing them for you. That's already useful — but it has a ceiling. Ask Claude something that spans ten notes and it has to open them one by one, hunting for the answer. That's not intelligence, that's grep with extra steps. The fix isn't more notes. It's connecting the ones you have into a graph — so your AI stops searching file by file and starts traversing the relationships between them.
Why flat notes hit a ceiling
A folder full of notes is a pile, even a well-organized one. Claude can read any single note instantly, but "how does this idea connect to that one?" means checking every file for a match — slow, and it misses connections that aren't spelled out in either note. A graph fixes this by making the relationships themselves something Claude can read directly, instead of something it has to rediscover every time you ask.
The same notes, first as a flat pile and then morphing into a connected web — same notes, the leap is the links:
Vault — flat notes
Same notes — the leap is the links.
Step 1 — Make the links explicit
A graph needs raw material: links that actually exist, not just implied ones. If you set up RULES.md in episode 01, you're already halfway there — every note connects with [[wikilinks]] and has tags in its frontmatter. If some of your notes are still sitting there unlinked, do one pass:
Read every note in my vault. Wherever two notes are actually related,add a [[wikilink]] between them and make sure both have matching tagsin their frontmatter. Don't force connections that aren't real.
That's the raw material — links and tags are just edges waiting to be read.
Step 2 — Have Claude build the map
Now turn those links into something Claude can query directly, instead of re-deriving every time. Paste this:
Read every note in my vault and build a knowledge graph of it ingraph/index.md:- One line per note: title, one-line summary, and what it links to (via [[wikilinks]] or shared tags)- Group notes that share a tag under a headingKeep this file updated whenever notes are added or linked.
That's it. graph/index.md becomes a map Claude reads before it answers anything relational — nodes are your notes, edges are the links and shared tags between them. It's just structured text, no tool required, and Claude maintains it the same way it maintains index.md from episode 01.
Step 3 — Query by relationship, not keyword
This is the actual payoff. Instead of asking Claude to find a note, ask it to walk the graph:
"How does my note on agents connect to the one on context windows?" — Claude follows the links between them instead of reading both blind.
"What relates to [[MCP]]?" — it walks every edge touching that note, not just files with "MCP" in the text.
"Find notes that are connected to a lot of others but have no summary yet." — the kind of question you can only ask once relationships are structured data.
None of these are keyword search. They're traversal — the same shift a database gets from a table scan to an index.
Keep the map growing
Same habit as episode 01, one step further: after you tell Claude "save what we just did to my vault," also let it update graph/index.md with the new note's links. Every note you add makes the whole map more useful, not just one file longer.
What you actually get out of this
Once the graph exists, Claude stops answering from one note and starts reasoning across your whole knowledge — it can spot a pattern connecting something you wrote in January to something from last week, because the link is right there in the map, not buried in your memory. That's the real difference between a pile of notes and a second brain: one you have to search, the other one thinks with you.
No special software, no database to run — just links you already made, a text file Claude keeps current, and questions shaped like "how does this connect" instead of "find me this."
AI FROM MY POV