💬 Ask Claude Code using the graph

Claude Code only uses what sees. If your second brain — the vault generated by Graphify—it’s tucked away in a corner of the disk that it never checks, so it’s as if it didn’t exist. The first step is always to put the vault inside (or next to) the project and tell the agent where it is, right at the start of the session.

It’s not enough for the vault to be on disk: the agent needs to be directed it before you start scanning the repository. The simplest way is an opening sentence that establishes the vault as primary source. The block below is a ready-to-paste prompt.

O segundo cérebro deste projeto está em graph-imports/<nome-do-grafo>/ .
Lá tem um .md por conceito (com backlinks [[...]]) e o GRAPH_REPORT.md
com os god nodes. Use ESSE vault como fonte primária antes de varrer o repo.

Where to start a question? With the god nodes — the most connected entities in the graph, listed in the GRAPH_REPORT.md. They are the hubs: entering through one of them brings the rest of the map along, because almost everything passes through there. Starting at a peripheral node is like entering through the back door.

It came in through a god node—and now what? The trick is to let the agent follow the backlinks: the [[links]] that each note carries to neighboring concepts. Instead of stopping at the first node, it follows the neighborhood and builds an answer connected, not isolated.

That’s exactly what the vault is designed to allow: each .md points to the neighbors, so the agent navigates the meaning — not loose text positions. Visually, the path looks like this:

↑ The question comes in through god node highlighted; from there, the agent follows the backlinks all the way to the neighbors (cyan) and returns an answer that connects the concepts—not an isolated paragraph.

All of this becomes a habit with a prompt well put together. The difference between "explain X" and a prompt that points to the graph is huge: the second one tells the agent to enter through the hub, follow backlinks, cite the notes, and stay within the scope from the vault. The block below is this module’s main copy-run.

Use o vault em graph-imports/<nome-do-grafo>/ para explicar <tema>.
Comece pelo god node mais ligado a ele, siga os backlinks [[...]] até os
conceitos vizinhos e cite cada nota .md que usar. Não varra o resto do repo.

Reading Markdown files works — but Graphify can also serve the graph as tools programmatically, via MCP. Then the agent queries the graph precisely (a node’s neighbors, the shortest path) instead of relying only on the text. It’s optional, but powerful for large graphs.

# dentro do Claude Code (skill form):
/graphify . --mcp

# ou via terminal puro:
graphify --mcp

# tools expostas: query_graph · get_node · get_neighbors · shortest_path

Trust, but verify. The graph stores the provenance for each node—the source file the entity came from—inside the graph.json. When the agent makes a claim, you close the loop by tracing it from the concept to the source doc original and checking that it matches. It's the best defense against hallucination.

↑ The stated concept leads, through provenance of the graph.json, all the way to the original source doc, where you can check. The source docs don’t go into the vault: they stay in the project.