Graphify (Apache-2.0, local tree-sitter parsing) pre-builds a knowledge graph of your project folder so Grok doesn't waste tokens re-scanning every file per session. Original benchmark: 71.5x fewer tokens vs raw file reads on mixed corpus.
Workflow:
1. Point Graphify at project root → parses code locally, extracts notes/PDFs
2. Output: GRAPH_REPORT.md (high-degree nodes, clusters, non-obvious edges)
3. Drop report into Grok Bot's standing brief with instruction: "read graph before opening files"
4. Bot now queries graph first → opens only relevant cluster files → fails predictably if graph is stale
Rebuild trigger: file hash change. For repos under 100 files, speedup is marginal. For actual working vaults, it eliminates the "rebuild the room from scratch" tax every morning.
Bonus: Same pattern works with Gmail archive as knowledge graph (subject = taxonomy, body = ontology). Graphify is the folder version of that 2004-era email scrapbook system.
Practical win isn't the 5-10x speedup on structural queries—it's forcing the model to check the card catalog before wandering the shelves.
Workflow:
1. Point Graphify at project root → parses code locally, extracts notes/PDFs
2. Output: GRAPH_REPORT.md (high-degree nodes, clusters, non-obvious edges)
3. Drop report into Grok Bot's standing brief with instruction: "read graph before opening files"
4. Bot now queries graph first → opens only relevant cluster files → fails predictably if graph is stale
Rebuild trigger: file hash change. For repos under 100 files, speedup is marginal. For actual working vaults, it eliminates the "rebuild the room from scratch" tax every morning.
Bonus: Same pattern works with Gmail archive as knowledge graph (subject = taxonomy, body = ontology). Graphify is the folder version of that 2004-era email scrapbook system.
Practical win isn't the 5-10x speedup on structural queries—it's forcing the model to check the card catalog before wandering the shelves.