How I Give a Grok Bot the Map Before the Files Graphify and Obsidian turn a project folder into a knowledge graph so the model stops rereading every file at the start of a session. Graphify is open source, Apache-2.0, and it runs the code parse on the machine. Obsidian is free for personal and commercial use, and the vault is ordinary markdown you keep. Graphify’s own material cites 71.5 times fewer tokens against reading the raw files, on one mixed corpus. I have not run that benchmark. A Grok Bot is a standing brief, a conversation that does not die when I close the tab, and a computer: browser, files, terminal. Left alone, it opens the folder, reads what it can reach, and rebuilds the room from scratch. On a small repo that is cheap. On the vault I actually work from, it is the same tax every morning. I have kept the other half of this system since the week Gmail shipped, April 2004. Each clipping is an email. The subject line is the taxonomy. The body carries the ontology: what the thing is, what it sits next to, what it is not. When Grok can read the account, the scrapbook is already a graph. Graphify is the same idea pointed at a folder instead of an inbox. I point it at the project the Bot is hired to keep. Code is parsed locally, with tree-sitter, and nothing in that pass has to leave the machine. Notes, PDFs, and the decision memos go through an extraction pass. The output I care about is not the interactive HTML. It is GRAPH_REPORT.md and the graph file beside it: the high-degree nodes, the clusters, the edges that are not obvious from the filenames. In Obsidian can read that map. The Bot does not need the canvas. It needs the report, and a line in its brief that says read the report before you open a file. Graphify’s installer writes that instruction into AGENTS.md. For a Grok Bot I put the same sentence in the standing brief and I leave the report in the Bot’s files. The first question of a session is not what is in this repo. It is what the graph says this touches. Then it opens the few files in that cluster. If the graph is wrong, the Bot is wrong in a way I can see, because the miss has an edge I can check. People who have timed it on pure code put the realistic cut closer to five or ten times on structural questions, and near nothing under a hundred files. I treat that number as someone else’s measurement. What I keep is the order of operations. Build the map once. Make the hire read the map. Rebuild when the hash of a file changes. Grok Bot is the employee that used to walk the shelves. Now it reads the card catalog first.